Try the Agent API Playground
Test Agent API requests and parameters interactively in the API console.
Pricing
Pay-as-you-go pricing for all APIs. No subscription required.
Why Use the Agent API?
Web-Grounded Answers
Get accurate, up-to-date answers grounded in real-time web search, with inline citations in a single call, and conversation context across turns.
Multi-Provider Access
Access OpenAI, Anthropic, Google, xAI, and more through one unified API, no need to manage multiple API keys.
Transparent Pricing
See exact token counts and costs per request, no markup, just direct provider pricing.
Granular Control
Change models, reasoning, tokens, and tools with consistent syntax.
We recommend using our official SDKs for a more convenient and type-safe way to interact with the Agent API.
Endpoint: The Agent API is available at
POST https://api.perplexity.ai/v1/agent. For OpenAI SDK compatibility, POST /v1/responses is also accepted as an alias. See the OpenAI Compatibility Guide for details on using OpenAI SDKs with Perplexity.Installation
Install the SDK for your preferred language:pip install perplexityai
npm install @perplexity-ai/perplexity_ai
Authentication
Set your API key as an environment variable. The SDK will automatically read it:- macOS/Linux
- Windows
export PERPLEXITY_API_KEY="your_api_key_here"
setx PERPLEXITY_API_KEY "your_api_key_here"
All SDK examples below automatically use the
PERPLEXITY_API_KEY environment variable. You can also pass the key explicitly if needed.Basic Usage
Convenience Property: Both Python and Typescript SDKs provide an
output_text property that aggregates all text content from response outputs. Instead of iterating through response.output, simply use response.output_text for cleaner code.Using a Third-Party Model
Use third-party models from OpenAI, Anthropic, Google, xAI, and other providers for specific capabilities:from perplexity import Perplexity
client = Perplexity()
response = client.responses.create(
model="openai/gpt-5.6-sol",
input="Explain the difference between supervised and unsupervised learning in machine learning."
)
print(f"Response ID: {response.id}")
print(response.output_text)
import Perplexity from '@perplexity-ai/perplexity_ai';
const client = new Perplexity();
const response = await client.responses.create({
model: "openai/gpt-5.6-sol",
input: "Explain the difference between supervised and unsupervised learning in machine learning."
});
console.log(`Response ID: ${response.id}`);
console.log(response.output_text);
curl https://api.perplexity.ai/v1/agent \
-H "Authorization: Bearer $PERPLEXITY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.6-sol",
"input": "Explain the difference between supervised and unsupervised learning in machine learning."
}' | jq
Response
Response
{
"background": false,
"completed_at": 1771891464,
"created_at": 1771891464,
"error": null,
"frequency_penalty": 0,
"id": "resp_f854ed0a-f0e2-4ee8-b5ea-8582956910f2",
"incomplete_details": null,
"instructions": null,
"max_output_tokens": null,
"max_tool_calls": null,
"metadata": {},
"model": "openai/gpt-5.6-sol",
"object": "response",
"output": [
{
"content": [
{
"annotations": [],
"logprobs": [],
"text": "Supervised learning uses labeled data where each example has a known output, enabling the model to learn direct input-output relationships. Examples include classification and regression.",
"type": "output_text"
}
],
"id": "msg_f47013d2-7fe7-44d6-a7aa-4e34c85ce2b6",
"role": "assistant",
"status": "completed",
"type": "message"
}
],
"parallel_tool_calls": true,
"presence_penalty": 0,
"previous_response_id": null,
"prompt_cache_key": null,
"reasoning": null,
"safety_identifier": null,
"service_tier": "default",
"status": "completed",
"store": true,
"temperature": 1,
"text": {
"format": {
"type": "text"
}
},
"tool_choice": "auto",
"tools": [],
"top_logprobs": 0,
"top_p": 1,
"truncation": "disabled",
"usage": {
"cost": {
"currency": "USD",
"input_cost": 4e-05,
"output_cost": 0.00311,
"total_cost": 0.00315
},
"input_tokens": 20,
"input_tokens_details": {
"cached_tokens": 0
},
"output_tokens": 222,
"output_tokens_details": {
"reasoning_tokens": 0
},
"total_tokens": 242
},
"user": null
}
Using a Preset
Presets provide optimized defaults for specific use cases. Start with a preset for quick setup:from perplexity import Perplexity
client = Perplexity()
response = client.responses.create(
preset="low",
input="Explain what the MMLU benchmark measures for large language models, and how the Apache 2.0 license differs from a restricted research-only license for model weights.",
)
print(f"Model used: {response.model}")
print(response.output_text)
import Perplexity from '@perplexity-ai/perplexity_ai';
const client = new Perplexity();
const response = await client.responses.create({
preset: "low",
input: "Explain what the MMLU benchmark measures for large language models, and how the Apache 2.0 license differs from a restricted research-only license for model weights.",
});
console.log(`Model used: ${response.model}`);
console.log(response.output_text);
curl https://api.perplexity.ai/v1/agent \
-H "Authorization: Bearer $PERPLEXITY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"preset": "low",
"input": "Explain what the MMLU benchmark measures for large language models, and how the Apache 2.0 license differs from a restricted research-only license for model weights."
}' | jq
Response
Response
{
"background": false,
"completed_at": 1771891641,
"created_at": 1771891641,
"error": null,
"frequency_penalty": 0,
"id": "resp_aca2bace-3782-4d81-be45-a82c24cfff9d",
"incomplete_details": null,
"instructions": "## Abstract\n<role>\nYou are an AI assistant developed by Perplexity AI...\n</role>\n...",
"max_output_tokens": 8192,
"max_tool_calls": null,
"metadata": {},
"model": "openai/gpt-5.1",
"object": "response",
"output": [
{
"queries": [
"2025 open source LLM benchmark performance",
"2025 newly released open source LLMs license",
"2025 open source LLM real world use cases"
],
"results": [
{
"date": "2025-11-19",
"id": 1,
"last_updated": "2026-02-23T12:12:34",
"snippet": "updated\n\n19 Nov 2025\n\n# Open LLM Leaderboard\n\nThis LLM leaderboard displays...",
"source": "web",
"title": "Open LLM Leaderboard 2025",
"url": "https://www.vellum.ai/open-llm-leaderboard"
},
{
"date": "2023-05-05",
"id": 2,
"last_updated": "2026-01-06T09:02:43.651546",
"snippet": "",
"source": "web",
"title": "A list of open LLMs available for commercial use.",
"url": "https://github.com/eugeneyan/open-llms"
},
{
"date": "2025-05-05",
"id": 3,
"last_updated": "2026-02-22T19:27:06",
"snippet": "# Best Open Source LLMs You Can Run Locally in 2025\n\nRunning large language models on your own hardware is...",
"source": "web",
"title": "Best Open Source LLMs You Can Run Locally in 2025 - DemoDazzle",
"url": "https://demodazzle.com/blog/open-source-llms-2025"
},
{
"date": "2025-12-15",
"id": 4,
"last_updated": "2026-02-23T21:56:51",
"snippet": "updated\n\n15 Dec 2025\n\n# LLM Leaderboard\n\nThis LLM leaderboard displays the latest public benchmark performance for SOTA model versions released after April 2024...",
"source": "web",
"title": "LLM Leaderboard 2025 - Vellum",
"url": "https://www.vellum.ai/llm-leaderboard"
},
{
"date": "2025-11-22",
"id": 5,
"last_updated": "2026-02-11T02:35:36",
"snippet": "Open\u2011source Large Language Models (LLMs) have moved from niche hobby projects to a full\u2011blown industry trend in 2025...",
"source": "web",
"title": "Open\u2011Source LLMs 2025: GPT\u2011OSS Models & How ... - Neura AI Blog",
"url": "https://blog.meetneura.ai/open-source-llms-2025/"
},
{
"date": "2025-07-23",
"id": 6,
"last_updated": "2026-02-23T23:43:21",
"snippet": "",
"source": "web",
"title": "55 real-world LLM applications and use cases from top ...",
"url": "https://www.evidentlyai.com/blog/llm-applications"
},
{
"date": "2025-10-29",
"id": 7,
"last_updated": "2026-02-23T21:22:10",
"snippet": "",
"source": "web",
"title": "Top 10 open source LLMs for 2025 - NetApp Instaclustr",
"url": "https://www.instaclustr.com/education/open-source-ai/top-10-open-source-llms-for-2025/"
},
{
"date": "2025-05-21",
"id": 8,
"last_updated": "2026-02-23T14:54:20",
"snippet": "Here are the details of OpenLLaMA:\n\n**Parameters:** 3B, 7B and 13B\n\n**License:** Apache 2.0...",
"source": "web",
"title": "The List of 11 Most Popular Open Source LLMs [2025]",
"url": "https://www.lakera.ai/blog/open-source-llms"
},
{
"date": "2026-01-07",
"id": 9,
"last_updated": "2026-02-23T17:41:06",
"snippet": "",
"source": "web",
"title": "The state of open source AI models in 2025 | Red Hat Developer",
"url": "https://developers.redhat.com/articles/2026/01/07/state-open-source-ai-models-2025"
},
{
"date": "2025-10-28",
"id": 10,
"last_updated": "2026-02-23T07:53:56",
"snippet": "- **Open source dominates by volume:** 63% of models in our dataset (59 open source vs 35 proprietary)\n- **Performance...",
"source": "web",
"title": "Open Source vs Proprietary LLMs: Complete 2025 Benchmark ...",
"url": "https://whatllm.org/blog/open-source-vs-proprietary-llms-2025"
},
{
"date": "2025-06-02",
"id": 11,
"last_updated": "2026-01-18T13:27:38.757741",
"snippet": "",
"source": "web",
"title": "Top 8 Open\u2011Source LLMs to Watch in 2025 - JetRuby Agency",
"url": "https://jetruby.com/blog/top-8-open-source-llms-to-watch-in-2025/"
},
{
"date": "2026-01-26",
"id": 12,
"last_updated": "2026-02-23T16:49:21",
"snippet": "",
"source": "web",
"title": "Best Open Source LLMs in 2026",
"url": "https://www.keywordsai.co/blog/best-open-source-llms"
},
{
"date": "2025-12-10",
"id": 13,
"last_updated": "2026-02-23T18:38:26",
"snippet": "",
"source": "web",
"title": "Full Benchmark Table For...",
"url": "https://skywork.ai/blog/llm/top-10-open-llms-2025-november-ranking-analysis/"
},
{
"date": "2024-09-19",
"id": 14,
"last_updated": "2025-12-27T09:28:04.559969",
"snippet": "## Top Open-Source LLMs of 2025\n\n### 1. LLaMA 3.1\n\n**Developer:**Meta AI **Release Date:**July 23, 2024 **Parameter Size:**405B, 70B, 8B...",
"source": "web",
"title": "Top 10 Open-Source LLMs in 2025 - Kite Metric",
"url": "https://kitemetric.com/blogs/top-10-open-source-llms-in-2025-a-comprehensive-guide"
},
{
"date": "2025-02-26",
"id": 15,
"last_updated": "2025-09-10T16:36:09.704235",
"snippet": "Use Cases:\n\n**Advanced Chatbots:**Responsive customer support bots. **Content Creation for Marketing:**Generating product descriptions and blog posts...",
"source": "web",
"title": "Top 10 Open-Source LLMs in 2025 and Their Use Cases",
"url": "https://capalearning.com/2025/02/26/top-10-open-source-llms-in-2025-and-their-use-cases/"
}
],
"type": "search_results"
},
{
"contents": [
{
"snippet": "Hi, Camille\u2019s here! On October 28, 2025, I fell into a small rabbit hole...",
"title": "Full Benchmark Table For...",
"url": "https://skywork.ai/blog/llm/top-10-open-llms-2025-november-ranking-analysis/"
},
{
"snippet": "# Open source vs proprietary LLMs: complete 2025 benchmark analysis\n\n## TL;DR: The state of LLMs in late 2025\n\n**The landscape has shifted dramatically:**\n\n- **Open source dominates by volume:** 63% of models in our dataset (59 open source vs 35 proprietary)\n- **Performance...",
"title": "Open Source vs Proprietary LLMs: Complete 2025 Benchmark ...",
"url": "https://whatllm.org/blog/open-source-vs-proprietary-llms-2025"
}
],
"type": "fetch_url_results"
},
{
"content": [
{
"annotations": [],
"logprobs": [],
"text": "In 2025, the strongest open\u2011source LLMs (Qwen 2.5, Llama 3.3/3.x, DeepSeek V3\u2011series, Mixtral...",
"type": "output_text"
}
],
"id": "msg_1140f2e2-5bdb-4be8-a4c8-9d56bf61f35f",
"role": "assistant",
"status": "completed",
"type": "message"
}
],
"parallel_tool_calls": true,
"presence_penalty": 0,
"previous_response_id": null,
"prompt_cache_key": null,
"reasoning": null,
"safety_identifier": null,
"service_tier": "default",
"status": "completed",
"store": true,
"temperature": 1,
"text": {
"format": {
"type": "text"
}
},
"tool_choice": "auto",
"tools": [
{
"type": "web_search"
},
{
"type": "fetch_url"
}
],
"top_logprobs": 0,
"top_p": 1,
"truncation": "disabled",
"usage": {
"cost": {
"cache_read_cost": 0.00059,
"currency": "USD",
"input_cost": 0.00919,
"output_cost": 0.02743,
"tool_calls_cost": 0.0055,
"total_cost": 0.04271
},
"input_tokens": 12088,
"input_tokens_details": {
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 4736,
"cached_tokens": 4736
},
"output_tokens": 2743,
"output_tokens_details": {
"reasoning_tokens": 0
},
"tool_calls_details": {
"fetch_url": {
"invocation": 1
},
"search_web": {
"invocation": 1
}
},
"total_tokens": 14831
},
"user": null
}
Learn more about presets to explore pre-configured setups optimized for different use cases with specific models, token limits, and tool access.
With Web Search
The Agent API provides access to a number of tools that can be used to extend the capabilities of the model. Enable web search capabilities using theweb_search tool:
from perplexity import Perplexity
client = Perplexity()
response = client.responses.create(
model="openai/gpt-5.6-sol",
input="Explain the original Transformer architecture from 'Attention Is All You Need' (Vaswani et al. 2017): encoder-decoder structure, multi-head self-attention, and positional encodings.",
tools=[{"type": "web_search"}],
instructions="You have access to a web_search tool. Use it for questions about current events, news, or recent developments. Use 1 query for simple questions. Keep queries brief: 2-5 words. NEVER ask permission to search - just search when appropriate",
)
if response.status == "completed":
print(response.output_text)
import Perplexity from '@perplexity-ai/perplexity_ai';
const client = new Perplexity();
const response = await client.responses.create({
model: "openai/gpt-5.6-sol",
input: "Explain the original Transformer architecture from 'Attention Is All You Need' (Vaswani et al. 2017): encoder-decoder structure, multi-head self-attention, and positional encodings.",
tools: [{ type: "web_search" }],
instructions: "You have access to a web_search tool. Use it for questions about current events, news, or recent developments.",
});
if (response.status === "completed") {
console.log(response.output_text);
}
curl https://api.perplexity.ai/v1/agent \
-H "Authorization: Bearer $PERPLEXITY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.6-sol",
"input": "Explain the original Transformer architecture from 'Attention Is All You Need' (Vaswani et al. 2017): encoder-decoder structure, multi-head self-attention, and positional encodings.",
"tools": [{"type": "web_search"}],
"instructions": "You have access to a web_search tool. Use it for questions about current events, news, or recent developments."
}' | jq
Response
Response
{
"background": false,
"completed_at": 1771891737,
"created_at": 1771891737,
"error": null,
"frequency_penalty": 0,
"id": "resp_367113ed-7a1b-4b2e-bad7-93e53a6cbeca",
"incomplete_details": null,
"instructions": "You have access to a web_search tool. Use it for questions about current events, news, or recent developments. Use 1 query for simple questions. Keep queries brief: 2-5 words. NEVER ask permission to search - just search when appropriate",
"max_output_tokens": 8192,
"max_tool_calls": null,
"metadata": {},
"model": "openai/gpt-5.6-sol",
"object": "response",
"output": [
{
"queries": [
"latest AI developments 2026"
],
"results": [
{
"date": "2026-01-01",
"id": 1,
"last_updated": "2026-02-23T20:10:25",
"snippet": "Many believe efficiency will be the new frontier...",
"source": "web",
"title": "The trends that will shape AI and tech in 2026 - IBM",
"url": "https://www.ibm.com/think/news/ai-tech-trends-predictions-2026"
},
{
"date": "2026-01-08",
"id": 2,
"last_updated": "2026-02-23T20:19:20",
"snippet": "## What\u2019s next in AI: 7 trends to watch in 2026\n\nAI is entering a new phase, one defined by real-world impact...",
"source": "web",
"title": "What's next in AI: 7 trends to watch in 2026 - Microsoft Source",
"url": "https://news.microsoft.com/source/features/ai/whats-next-in-ai-7-trends-to-watch-in-2026/"
},
{
"date": "2026-01-06",
"id": 3,
"last_updated": "2026-02-21T02:30:13",
"snippet": "#### Topics\n\n#### AI in Action\n\n**Summary:**\n\nMIT SMR columnists Thomas H. Davenport and Randy Bean see five...",
"source": "web",
"title": "Five Trends in AI and Data Science for 2026",
"url": "https://sloanreview.mit.edu/article/five-trends-in-ai-and-data-science-for-2026/"
},
{
"date": "2026-01-06",
"id": 4,
"last_updated": "2026-02-24T00:01:21",
"snippet": "## Jeff Su\n\n##### Jan 06, 2026 (0:13:13)\nMost #AI predictions are speculation. This video covers...",
"source": "web",
"title": "Top 6 AI Trends That Will Define 2026 (backed by data)",
"url": "https://www.youtube.com/watch?v=B23W1gRT9eY"
},
{
"date": "2026-01-15",
"id": 5,
"last_updated": "2026-02-23T17:37:52",
"snippet": "",
"source": "web",
"title": "11 things AI experts are watching for in 2026 | University of California",
"url": "https://www.universityofcalifornia.edu/news/11-things-ai-experts-are-watching-2026"
},
{
"date": "2026-01-13",
"id": 6,
"last_updated": "2026-02-23T16:27:23",
"snippet": "Artificial intelligence (AI) is no longer an emerging technology, it\u2019s a transformational force driving innovation across industries...",
"source": "web",
"title": "AI Trends in 2026: A New Era of AI Advancements and Breakthroughs",
"url": "https://www.trigyn.com/insights/ai-trends-2026-new-era-ai-advancements-and-breakthroughs"
},
{
"date": "2025-12-22",
"id": 7,
"last_updated": "2026-02-23T09:47:25",
"snippet": "The most significant advances in artificial intelligence next year won't come from...",
"source": "web",
"title": "6 AI breakthroughs that will define 2026 - InfoWorld",
"url": "https://www.infoworld.com/article/4108092/6-ai-breakthroughs-that-will-define-2026.html"
},
{
"date": "2025-12-22",
"id": 8,
"last_updated": "2026-02-23T20:21:57",
"snippet": "What will define AI in 2026? \ud83d\ude80 Martin Keen & Aaron Baughman explore groundbreaking trends like Agentic AI, cloud computing, automation, and quantum computing, plus innovations like Physical AI...",
"source": "web",
"title": "AI Trends 2026: Quantum, Agentic AI & Smarter Automation",
"url": "https://www.youtube.com/watch?v=zt0JA5rxdfM"
},
{
"date": "2025-12-15",
"id": 9,
"last_updated": "2026-02-23T13:13:58",
"snippet": "",
"source": "web",
"title": "Stanford AI Experts Predict What Will Happen in 2026",
"url": "https://hai.stanford.edu/news/stanford-ai-experts-predict-what-will-happen-in-2026"
},
{
"date": "2025-05-10",
"id": 10,
"last_updated": "2026-02-20T16:07:11",
"snippet": "{ts:574} breakthroughs in AlphaGo and Alpha Fold, which are absolutely incredible. Now, DeepMind has basically said...",
"title": "2026 AI : 10 Things Coming In 2026 (A.I In 2026 Major Predictions)",
"url": "https://www.youtube.com/watch?v=RfA2Ug4FuaY"
}
],
"type": "search_results"
},
{
"content": [
{
"annotations": [],
"logprobs": [],
"text": "Here are major *recent* directions in AI (late 2025\u2013early 2026) that researchers...",
"type": "output_text"
}
],
"id": "msg_d0f12cc6-c6a2-426f-b55e-fff247e40c8c",
"role": "assistant",
"status": "completed",
"type": "message"
}
],
"parallel_tool_calls": true,
"presence_penalty": 0,
"previous_response_id": null,
"prompt_cache_key": null,
"reasoning": null,
"safety_identifier": null,
"service_tier": "default",
"status": "completed",
"store": true,
"temperature": 1,
"text": {
"format": {
"type": "text"
}
},
"tool_choice": "auto",
"tools": [
{
"type": "web_search"
}
],
"top_logprobs": 0,
"top_p": 1,
"truncation": "disabled",
"usage": {
"cost": {
"currency": "USD",
"input_cost": 0.00826,
"output_cost": 0.0063,
"tool_calls_cost": 0.005,
"total_cost": 0.01956
},
"input_tokens": 4718,
"input_tokens_details": {
"cached_tokens": 0
},
"output_tokens": 450,
"output_tokens_details": {
"reasoning_tokens": 0
},
"tool_calls_details": {
"search_web": {
"invocation": 1
}
},
"total_tokens": 5168
},
"user": null
}
With Finance Search
Retrieve structured financial and market data using thefinance_search tool. See the Finance Search guide for capabilities and recommended configurations.
from perplexity import Perplexity
client = Perplexity()
response = client.responses.create(
model="openai/gpt-5.6-sol",
input="Explain how to read a 10-K filing: what each major section contains (Item 1 Business, Item 1A Risk Factors, Item 7 MD&A, Item 8 Financial Statements) and how investors use them.",
tools=[{"type": "finance_search"}],
)
for item in response.output:
if item.type == "message":
print(item.content[0].text)
import Perplexity from '@perplexity-ai/perplexity_ai';
const client = new Perplexity();
const response = await client.responses.create({
model: "openai/gpt-5.6-sol",
input: "Explain how to read a 10-K filing: what each major section contains (Item 1 Business, Item 1A Risk Factors, Item 7 MD&A, Item 8 Financial Statements) and how investors use them.",
tools: [{ type: "finance_search" }],
});
console.log(response.output_text);
curl https://api.perplexity.ai/v1/agent \
-H "Authorization: Bearer $PERPLEXITY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.6-sol",
"input": "Explain NVIDIA's GPU compute model: streaming multiprocessors, CUDA cores, Tensor Cores, and HBM memory bandwidth.",
"tools": [{"type": "finance_search"}]
}' | jq
Response — Explain how to read a 10-K filing: what each major section contains (Item 1 Business, I...
Response — Explain how to read a 10-K filing: what each major section contains (Item 1 Business, I...
{
"id": "resp_4c946a0e-9a51-44d1-89a6-c972c57228a8",
"created_at": 1779391718,
"model": "openai/gpt-5.1",
"object": "response",
"output": [
{
"results": [
{
"id": 1,
"snippet": "(d) In response to Item l, Business, such registrant only need furnish a brief\ndescription of the business done by the registrant and its subsidiaries during the\nmost recent fiscal year which will, in the opinion of management, indicate the\ngeneral nature and scope of the business of the registrant and its subsidiaries, and\nin response to Item 2, Properties, such registrant only need furnish a brief\ndescription of the material properties of the registrant and its subsidiaries to the\nextent, in the opinion of the management, necessary to an understanding of the\nbusiness done by the registrant and its subsidiaries.\n...\nfollowing otherwise required Items:\n(a) Item 1, Business;\n...\nPART I \n[See General Instruction G(2)] \nItem 1.\nBusiness.\nFurnish the information required by Item 101 of Regulation S-K (§ 229.101 of this chapter) \nexcept that the discussion of the development of the registrant’s business need only include\ndevelopments since the beginning of the fiscal year for which this report is filed.\nItem 1A.\nRisk Factors.\nSet forth, under the caption “Risk Factors,” where appropriate, the risk factors described in \nItem 105 of Regulation S-K (§ 229.105 of this chapter) applicable to the registrant.\nProvide any\ndiscussion of risk factors in plain English in accordance with Rule 421(d) of the Securities Act of \n1933 (§ 230.421(d) of this chapter).\nSmaller reporting companies are not required to provide the \ninformation required by this item.\nItem 1B.",
"title": "[PDF] Form 10-K - SEC.gov",
"url": "https://www.sec.gov/files/form10-k.pdf",
"date": null,
"last_updated": "2025-06-03",
"source": "web"
},
{
"id": 2,
"snippet": "Regulation S-K, Item 105, requires registrants to provide “a discussion of the\nmaterial factors that make an investment in the registrant or offering\nspeculative or risky.”\nCertain indicators of risk may be present in the\nfootnotes to the financial statements, in MD&A, or elsewhere in investor\npresentations or other periodic filings.",
"title": "3.3 Disclosures About Risk | DART – Deloitte Accounting Research ...",
"url": "https://dart.deloitte.com/USDART/home/publications/deloitte/additional-deloitte-guidance/roadmap-sec-comment-letter-considerations/chapter-3-sec-disclosure-topics/3-3-disclosures-about-risk",
"date": null,
"last_updated": "2026-03-31",
"source": "web"
},
{
"id": 3,
"snippet": "Additional sections in this Form 10-K which should be helpful to the reading of our discussion and analysis include the following: (i) a description of our services provided, by segment found in Items\n1 and 2 “Business and Properties”—”Services Provided” (ii) a description of our business strategy found in Items 1 and 2 “Business and Properties”—”Our Strategy”; and (iii) a description of\nrisk factors affecting us and our business, found in Item 1A “Risk Factors.”",
"title": "Form 10-K Item 7. Management's Discussion and Analysis - SEC.gov",
"url": "https://www.sec.gov/Archives/edgar/data/1449732/000119312512289206/d374099dex993.htm",
"date": "2012-03-29",
"last_updated": "2025-09-23",
"source": "web"
},
{
"id": 4,
"snippet": "In summary, Forms 10-K, 10-Q, 20-F and 40-F share detailed information and insights into the company’s overall financial performance and business operational details, while Forms 8-K and 6-K are filed to provide timely and relevant updates on significant material changes.",
"title": "How to navigate Forms 10-K, 10-Q, 20-F, 40-F, 8-K and 6-K",
"url": "https://www.toppanmerrill.com/blog/how-to-navigate-forms-10-k-10-q-20-f-40-f-8-k-and-6-k/",
"date": "2025-03-19",
"last_updated": "2026-05-16",
"source": "web"
},
{
"id": 5,
"snippet": "#### Item 1 – Business\nThis describes the business of the company: who and what the company does, what subsidiaries it owns, and what markets it operates in.\nIt may also include recent events, competition, regulations, and labor issues.\n(Some industries are heavily regulated, have complex labor requirements, which have significant effects on the business.)\nOther topics in this section may include special operating costs, seasonal factors, or insurance matters.",
"title": "Form 10-K - Wikipedia",
"url": "https://en.wikipedia.org/wiki/Form_10-K",
"date": "2005-02-23",
"last_updated": "2026-03-31",
"source": "web"
},
{
"id": 6,
"snippet": "A few companies located the summary in “Item 1.\nBusiness.”",
"title": "SEC Risk Factor Disclosure Rules",
"url": "https://corpgov.law.harvard.edu/2021/12/22/sec-risk-factor-disclosure-rules/",
"date": "2021-12-22",
"last_updated": "2026-04-13",
"source": "web"
},
{
"id": 7,
"snippet": "Regulation S-K, Item 303, specifies the information that a registrant is\nrequired to provide when discussing its financial condition and results of\noperations in MD&A.\n...\n- Requiring the disclosure of (1) any known trends or uncertainties that have had or are reasonably likely to have a material impact on revenues or income and (2) any known events that are “reasonably likely to cause a material change in the relationship between costs and revenues (such as known or reasonably likely future increases in costs . . . )” (emphasis added).\n...\nUnder Regulation S-K, Item 303, registrants are required to disclose in MD&A\nmaterial known trends or uncertainties that may affect future performance\n(whether favorable or unfavorable).\n...\nTo provide comprehensive and meaningful disclosures, management should consider disclosing the following items in the critical accounting policies section of MD&A:- The method(s) used to determine critical accounting estimates.\n- The accuracy of past estimates or assumptions.\n- The extent to which the estimates or assumptions have changed.\n- The drivers that affect variability.\n- Which estimates or assumptions are reasonably likely to change in the future.\n...\nmatters in MD&A if those matters meet the criteria of Regulation\nS-K, Item 303(b)(2)(ii), which requires disclosure of “any known trends\nor uncertainties that have had or that are reasonably likely to have” a\nmaterial impact on revenues or income.",
"title": "3.1 Management's Discussion and Analysis | DART",
"url": "https://dart.deloitte.com/USDART/home/publications/deloitte/additional-deloitte-guidance/roadmap-sec-comment-letter-considerations/chapter-3-sec-disclosure-topics/3-1-management-s-discussion-analysis",
"date": null,
"last_updated": "2026-04-19",
"source": "web"
},
{
"id": 8,
"snippet": "- **Item 1: Business**\n...\n## Item 1 - BusinessneCompanies typically define their business in this opening section of the 10-K report.\nThey describe their various product lines and business segments.\nThey list contracts, raw materials used, and supplier or distribution channels.\nThey talk about the competition and competitive factors in the market.\nIf research and development or intellectual property issues are important to company operations, they are included.\nGovernment regulations are covered.\nFinally, several pages are devoted to outlining risk factors to consider in evaluating the company's business.",
"title": "The 10-K - SEC Filings - Research Guides at Baruch College",
"url": "https://guides.newman.baruch.cuny.edu/c.php?g=188202&p=1244183",
"date": "2009-11-02",
"last_updated": "2026-05-09",
"source": "web"
},
{
"id": 9,
"snippet": "- **Item 1A: Risk Factors:** Absolutely critical reading.\nThis **section** lists the most significant risks and uncertainties that could materially affect the company’s business, **financial condition**, or operating results.\nEffective **10k risk factors analysis** is paramount for **investors**.\nLook for specific, quantifiable risks, not just boilerplate warnings.",
"title": "How to Read a 10-K Report with AI | Complete SEC Analysis Guide",
"url": "https://www.v7labs.com/blog/how-to-read-a-10k-report-ai-sec-filings-guide",
"date": "2025-06-11",
"last_updated": "2026-05-16",
"source": "web"
},
{
"id": 10,
"snippet": "**Item 7: Management’s Discussion and Analysis of Financial Condition and Results of Operations (MD&A)**– This section is perhaps the most narrative part of Form 10-K, where the company’s executives discuss the financial and operational factors that affected the business’s performance over the reporting period.\n...\nItem 7, Management’s Discussion and Analysis (MD&A), is an essential part of Form 10-K that offers investors a detailed narrative crafted by the company’s management.\nIt provides context and analysis beyond the figures presented in the financial statements.\nThis section aims to offer a view of the company through the lens of its management, explaining the dynamics of the business, the financial outcomes, and the strategies and decisions that influenced those results over the fiscal year.\n...\n**Operational Review**: This component of MD&A provides an analysis of the company’s business operations over the fiscal year.\nIt covers critical areas such as sales trends, customer acquisition and retention, changes in the competitive landscape, and operational milestones.",
"title": "What are Items 7, 7A, and 8 in Part II of Form 10-K? - Superfast CPA",
"url": "https://www.superfastcpa.com/what-are-items-7-7a-and-8-in-part-ii-of-form-10-k/",
"date": "2021-03-13",
"last_updated": "2026-01-05",
"source": "web"
},
{
"id": 11,
"snippet": "(a) If the registrant experiences a cybersecurity incident that is determined by the registrant \n...\nIf the registrant or any of its subsidiaries consolidated has completed the acquisition or \ndisposition of a significant amount of assets, otherwise than in the ordinary course of business, or \nthe acquisition or disposition of a significant amount of assets that constitute a real estate \noperation as defined in § 210.3-14(a)(2) disclose the following information:\n(a) the date of completion of the transaction; \n(b) a brief description of the assets involved; \n(c) the identity of the person(s) from whom the assets were acquired or to whom they were \nsold and the nature of any material relationship, other than in respect of the transaction, between \nsuch person(s) and the registrant or any of its affiliates, or any director or officer of the\n...\n(1) \nthe date on which the registrant becomes obligated on the direct financial obligation \nand a brief description of the transaction or agreement creating the obligation; \n(2)\nthe amount of the obligation, including the terms of its payment and, if applicable, a \nbrief description of the material terms under which it may be accelerated or increased and the \nnature of any recourse provisions that would enable the registrant to recover from third parties; \nand; \n...\na brief description of the other terms and conditions of the transaction or agreement\nthat are material to the registrant.",
"title": "[PDF] Form 8-K - SEC.gov",
"url": "https://www.sec.gov/files/form8-k.pdf",
"date": null,
"last_updated": "2025-04-11",
"source": "web"
},
{
"id": 12,
"snippet": "**Item 1 **“Business” requires a description of the company’s business, including its main products and services, what subsidiaries it owns, and what markets it operates in.\nThis section may also include information about recent events, competition the company faces, regulations that apply to it, labor issues, special operating costs, or seasonal factors.\nThis is a good place to start to understand how the company operates.",
"title": "How to Read a 10-K/10-Q | Investor.gov",
"url": "https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-bulletins/how-read",
"date": "2021-01-25",
"last_updated": "2026-05-17",
"source": "web"
},
{
"id": 13,
"snippet": "**Item 7: Management's Discussion and Analysis (MD&A)** is the required narrative section of the 10-K where management explains its financial results and financial condition.\nIt covers results of operations (why revenue and expenses changed), liquidity and capital resources (how the company funds itself), and critical accounting policies and estimates (the judgements that shaped the numbers).\nThe MD&A is not audited by the company's auditor; it is management's explanation of the statements.\nYet it is heavily scrutinised by the SEC, which requires it to be accurate, balanced, and not misleading — and has enforcement power to challenge inadequate disclosures.\n...\nThe SEC's MD&A requirements are detailed in Regulation S-K, Item 303.\nThe agency requires that MD&A address:\n1. **Results of operations.** A discussion of revenue, cost of goods sold, operating expenses, and operating income.\nThis must include year-over-year comparisons and explanation of significant changes (usually anything more than 5% variance).\n2. **Liquidity and capital resources.** Cash flow from operations, investing, and financing.\nManagement must discuss how it funds the business, what liquidity constraints exist, and what capital expenditures are planned.\n3. **Critical accounting policies and estimates.** The accounting methods and assumptions that have the most impact on financial results.\nExamples include revenue recognition, inventory valuation, allowances for doubtful accounts, and pension assumptions.\n4. **Off-balance-sheet arrangements and known contractual obligations.** Leases, purchase commitments, debt covenants, and other obligations that shape future cash requirements.\n5. **Tabular disclosure of contractual obligations.** A table showing debt maturity, operating lease obligations, purchase commitments, and other fixed obligations by period (current year, years 2–3, years 4–5, beyond 5 years).\n6. **Forward-looking statements and risk disclosures.** While Item 1A (Risk Factors) covers risks in depth, MD&A must touch on risks relevant to the company's financial condition and results.\n...\nItem 7 (MD&A) is the narrative bridge between the audited financial statements and the real business.",
"title": "Item 7: Management's Discussion and Analysis (MD&A)",
"url": "https://pomegra.io/learn/library/track-b-stock-market-core/financial-statements/chapter-08-reading-a-10-k/item-7-md-and-a",
"date": null,
"last_updated": "2026-05-14",
"source": "web"
}
],
"type": "search_results",
"queries": [
"SEC Form 10-K item 1 business description",
"SEC 10-K Item 1A risk factors explanation",
"SEC 10-K Item 7 MD&A overview",
"SEC 10-K Item 8 financial statements contents"
]
},
{
"id": "msg_e81fbad4-f4ac-4092-b332-d04db8f414b2",
"content": [
{
"text": "Investors read a Form 10‑K by focusing on a few core items that move from “what this business is” to “what could go wrong” to “how management explains the numbers” to the audited financials themselves.[web:12][web:5] Together, Items 1, 1A, 7, and 8 form the spine of the filing and should be read as a connected story, not as isolated sections.[web:1][web:13]\n\n## Item 1: Business\n\nItem 1 “Business” describes what the company actually does, including its main products and services, key subsidiaries, and the markets where it operates.[web:12][web:5] This section often discusses competition, applicable regulations, labor issues, important contracts, seasonal patterns, and sometimes special operating costs or insurance matters that affect operations.[web:12][web:8]\n\nInvestors use Item 1 to understand the company’s **economic** engine and competitive position, which is essential context for interpreting the rest of the 10‑K.[web:12][web:8] Common uses include:\n- Checking how the company makes money (revenue streams, segments, and geographies).[web:12][web:8]\n- Evaluating competitive landscape and regulatory exposure that may affect margins or growth.[web:8][web:5]\n- Seeing whether the described business model aligns with later financial performance and management’s narrative in MD&A.[web:3][web:13]\n\n## Item 1A: Risk Factors\n\nItem 1A “Risk Factors” must present the material factors that make an investment in the company speculative or risky, under Regulation S‑K Item 105.[web:1][web:2] It should be written in plain English and cover significant uncertainties that could materially affect the business, financial condition, or operating results.[web:1][web:9]\n\nInvestors use this section to assess downside scenarios and weed out boilerplate from company‑specific risks.[web:2][web:9] Typical approaches include:\n- Identifying the few most material, tailored risks (e.g., reliance on one product, tight liquidity, regulatory investigations) versus generic language.[web:2][web:9]\n- Mapping risks back to Item 1 and Item 7 to see whether management’s strategy and financial explanations realistically reflect those risks.[web:2][web:13]\n- Using the risk list as a checklist for stress‑testing valuation assumptions (growth, margins, leverage, and capital needs).[web:2][web:9]\n\n## Item 7: MD&A\n\nItem 7 “Management’s Discussion and Analysis of Financial Condition and Results of Operations” (MD&A) is the narrative where management explains the financial results and financial condition, as required by Regulation S‑K Item 303.[web:7][web:13] It must discuss results of operations, liquidity and capital resources, and critical accounting policies and estimates, including known trends or uncertainties reasonably likely to materially impact performance.[web:7][web:13]\n\nInvestors use MD&A as the **bridge** between the business description and the audited numbers.[web:10][web:13] Common uses include:\n- Understanding the drivers behind year‑over‑year changes in revenue, margins, and expenses, instead of just seeing the changes in the statements.[web:10][web:13]\n- Evaluating liquidity, funding sources, capital expenditures, and off‑balance‑sheet arrangements to judge solvency and future cash needs.[web:7][web:13]\n- Studying critical accounting estimates to spot areas where management judgment heavily affects reported earnings (e.g., revenue recognition, reserves, impairment).[web:7][web:13]\n\n## Item 8: Financial Statements\n\nItem 8 contains the audited financial statements and related notes, typically including the balance sheet, income statement, cash flow statement, statement of shareholders’ equity, and accompanying footnotes.[web:1][web:10] This section also includes the independent auditor’s report, which speaks to whether the financials are fairly presented in conformity with GAAP, and sometimes flags material weaknesses in internal control.[web:1][web:10]\n\nInvestors use Item 8 to test the narrative against hard numbers and to build models.[web:10][web:13] Key uses include:\n- Verifying that trends highlighted in MD&A actually show up in revenue, margins, cash flows, and leverage metrics.[web:10][web:13]\n- Mining footnotes for segment data, debt terms, contingencies, leases, and other detailed disclosures that may not be prominent elsewhere.[web:10][web:1]\n- Building or updating valuation models (DCF, comparables) using audited figures rather than management’s non‑GAAP metrics or selective highlights.[web:10][web:13]\n\n## How to read them together\n\nA practical way to read a 10‑K is to start with Item 1 (what the company is), move to Item 1A (what can go wrong), then Item 7 (how management explains recent performance and future trends), and finally Item 8 (whether the audited numbers and notes support that story).[web:12][web:13] Experienced investors constantly cross‑check among these sections, looking for inconsistencies between the business description, the risks, management’s discussion, and the underlying financial statements.[web:3][web:7]",
"type": "output_text",
"annotations": [],
"logprobs": []
}
],
"role": "assistant",
"status": "completed",
"type": "message"
}
],
"status": "completed",
"error": null,
"usage": {
"input_tokens": 7267,
"output_tokens": 1188,
"total_tokens": 8455,
"cost": {
"currency": "USD",
"input_cost": 0.0046,
"output_cost": 0.01188,
"total_cost": 0.02193,
"cache_creation_cost": null,
"cache_read_cost": 0.00045,
"tool_calls_cost": 0.005
},
"input_tokens_details": {
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 3584,
"cached_tokens": 3584
},
"tool_calls_details": {
"search_web": {
"invocation": 1
}
},
"output_tokens_details": {
"reasoning_tokens": 0
}
},
"background": false,
"completed_at": 1779391718,
"frequency_penalty": 0,
"incomplete_details": null,
"instructions": "## Abstract\n<role>\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n</role>\n\n## Instructions\n<tools_workflow>\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the <tool_call_limit> below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n<tool_call_limit> Make at most three tool calls before concluding.</tool_call_limit>\n</tools_workflow>\n\n## Citation Instructions\n<citation_instructions>\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n<common_source_types> are included below.\n\n<common_source_types>\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n</common_source_types>\n\n<formatting_citations>\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n</formatting_citations>\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n<response_guidelines>\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n<tone>\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n</tone>\n\n### Lists and Paragraphs\n<lists_and_paragraphs>\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n</lists_and_paragraphs>\n\n### Summaries and Conclusions\n<summaries_and_conclusions>\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n</summaries_and_conclusions>\n\n## Prohibited Meta-Commentary\n<prohibited_commentary>\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n</prohibited_commentary>\n\n<copyright_requirements>\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n</copyright_requirements>\n\nCurrent date: Thursday, May 21, 2026\n\n",
"max_output_tokens": 8192,
"max_tool_calls": null,
"metadata": {},
"parallel_tool_calls": true,
"presence_penalty": 0,
"previous_response_id": null,
"prompt_cache_key": null,
"reasoning": null,
"safety_identifier": null,
"service_tier": "default",
"store": true,
"temperature": 1,
"text": {
"format": {
"type": "text"
}
},
"tool_choice": "auto",
"tools": [
{
"type": "web_search"
},
{
"type": "fetch_url"
}
],
"top_logprobs": 0,
"top_p": 1,
"truncation": "disabled",
"user": null
}
Response — Explain NVIDIA's GPU compute model: streaming multiprocessors, CUDA cores, Tensor Cores...
Response — Explain NVIDIA's GPU compute model: streaming multiprocessors, CUDA cores, Tensor Cores...
{
"id": "resp_ef76b85e-1a63-4c0f-97ce-9e58dd1d050a",
"created_at": 1779391739,
"model": "openai/gpt-5.1",
"object": "response",
"output": [
{
"results": [
{
"id": 1,
"snippet": "When we program GPUs , we produce sequences of instructions for its Streaming Multiprocessors to carry out.\nStreaming Multiprocessors (SMs) of NVIDIA GPUs are roughly analogous to the\ncores of CPUs.\nThat is, SMs both execute computations and store state available\nfor computation in registers, with associated caches.\nCompared to CPU cores, GPU\nSMs are simple, weak processors.\nExecution in SMs is pipelined within an\ninstruction (as in almost all CPUs since the 1990s) but there is no speculative\nexecution or instruction pointer prediction (unlike all contemporary\nhigh-performance CPUs).\n...\nAn H100 SXM GPU draws at most 700 W and has 132 SMs, each of which has four\nWarp Schedulers that can each issue instructions to 32 threads (aka a warp ) in parallel per clock cycle, for a total of 128 × 132 > 16,000 parallel threads running at about 5 cW apiece.\n...\nGPU SMs also support a large number of *concurrent* threads -- threads of execution whose instructions are interleaved.\nA single SM on an H100 can concurrently execute up to 2048 threads split across\n64 thread groups of 32 threads each.\nWith 132 SMs, that's a total of over\n250,000 concurrent threads.\nCPUs can also run many threads concurrently.\nBut switches between\nwarps happen at the speed of a single clock cycle (over 1000x faster than context switches on a CPU), again powered by the SM's Warp Schedulers . The volume of available warps and the speed of warp switches help hide latency caused by memory reads, thread synchronization, or other expensive instructions, ensuring that the arithmetic bandwidth provided by the CUDA Cores and Tensor Cores is well utilized.",
"title": "What is a Streaming Multiprocessor? | GPU Glossary - Modal",
"url": "https://modal.com/gpu-glossary/device-hardware/streaming-multiprocessor",
"date": null,
"last_updated": "2026-05-15",
"source": "web"
},
{
"id": 2,
"snippet": "NVIDIA doesn’t call these tiles “cores” at all — it calls them Graphics Processing Clusters, or “GPCs”.\n...\nIn this area of the chip, we expect to find 16 load/store units, 4 special function units, 128 CUDA cores, and 4 Tensor cores.\n...\nStarting here on the GPU side, we know that each one of these streaming multiprocessors has 128 CUDA cores and 4 Tensor cores:\n...\nThis gives us a grand total of 8 Zen 4 cores on the CPU, 18,432 CUDA cores and 576 Tenser cores on the GPU:\n...\nSpecifically, we know that an Ada Streaming Multiprocessor has 128 CUDA cores and 4 Tensor cores.",
"title": "Zen, CUDA, and Tensor Cores, Part I: The Silicon",
"url": "https://www.computerenhance.com/p/zen-cuda-and-tensor-cores-part-i",
"date": "2024-09-03",
"last_updated": "2026-05-19",
"source": "web"
},
{
"id": 3,
"snippet": "The terminology section defines a Streaming Multiprocessor (SM) as something that: “executes compute instructions on the GPU.”\n...\nSMs are the GPU’s core units running compute instructions.\nGPU engines include SMs plus other parts like copy or video engines handling various tasks.",
"title": "Difference between Streaming Multiprocessor and Compute Engine?",
"url": "https://forums.developer.nvidia.com/t/difference-between-streaming-multiprocessor-and-compute-engine/300154",
"date": "2024-07-17",
"last_updated": "2026-04-28",
"source": "web"
},
{
"id": 4,
"snippet": "The INT32 units do integer calculations, the FP32 and FP64 units floating-point calculations.\nLD/ST are load-store units, the SFU calculates special functions (e.g. sin/cos).\n...\nAs @rs277 already explained, when people speak of a GPU with *n* “CUDA cores” they mean a GPU with *n* FP32 cores, each of which can perform one single-precision fused multiply-add operation (FMA) per cycle.\nThe number of “CUDA cores” does not indicate anything in particular about the number of 32-bit integer ALUs, or FP64 cores, or multi-function units, or “Tensor cores” (which I would also consider a marketing term).",
"title": "Understanding of Tensor Core, Cuda Core and other cores in ...",
"url": "https://forums.developer.nvidia.com/t/understanding-of-tensor-core-cuda-core-and-other-cores-in-ampere-architecture/235900",
"date": "2022-12-01",
"last_updated": "2026-05-10",
"source": "web"
},
{
"id": 5,
"snippet": "A Streaming Multiprocessor (SM) is a fundamental component of NVIDIA GPUs, consisting of multiple Stream Processors (CUDA Core) responsible for executing instructions in parallel.\nThey are general purpose processors with a low clock rate target and a small cache.\n...\nConsists of:\n- SUPER LARGE Register File - This is how they can context switch quickly with no overhead, by keeping data on registers, see Warp Scheduling\n- Caches and shared memory\n- Warp Scheduler\n- Execution units (SFUs, CUDA Cores and Tensor Cores)\n...\n> SMs execute several thread blocks in parallel.\nAs soon as one of its thread block has completed execution, it takes up the serially next thread block.\nFrom Stephen Jones, I learned that each SM can managed 64 warps, so a total of 2048 threads.\nHowever, it really processes 4 warps at a time (see Warp Scheduling).\n...\n> An SM may contain up to 8 thread blocks in total.\n...\n> In general, SMs support instruction-level parallelism but not branch prediction.\nEach architecture in GPU consists of several SM.",
"title": "Streaming Multiprocessor (SM) - Steven Gong",
"url": "https://stevengong.co/notes/Streaming-Multiprocessor",
"date": "2026-02-07",
"last_updated": "2026-05-20",
"source": "web"
},
{
"id": 6,
"snippet": "By tightly integrating these Tensor Cores with expanded special function units within NVIDIA Rubin’s streaming multiprocessors, the platform significantly accelerates attention mechanisms and sparse compute paths, boosting both arithmetic density and energy efficiency without compromising model accuracy.",
"title": "NVIDIA Tensor Cores",
"url": "https://www.nvidia.com/en-us/data-center/tensor-cores/",
"date": "2026-03-16",
"last_updated": "2026-05-20",
"source": "web"
},
{
"id": 7,
"snippet": "Nvidia solved the problem of escalating complexity with its \"unified\" Tesla architecture, released in 2006.\nIn the G80 die, there is no more distinction between layers.\nThe Stream Multiprocessor (SM) replaces all previous units thanks to its ability to run vertex, fragment and geometry \"kernel\" without distinction.\nThe load balancing happens automatically by swapping the \"kernel\" run by each SM depending on the need of the pipeline.\nNo longer SIMD capable, \"shaders units\" are now \"core\" capable of one integer or one float32 instruction per clock.\nSM receive threads in groups of 32 called warps.\nIdeally all threads in a warp will execute the same instruction at the same time, only on different data (hence the name SIMT).\nThe Multi-threaded Instruction Unit (MT) takes care of enabling/disabling threads in a warp in case their Instruction Pointer (IP) converge/diverge.\nTwo SFU units are here to help with complex mathematic calculation such as inverse square root, sin, cos, exp, and rcp.\nThese units are also able to execute one instruction per clock but since there are only two of them, warp execution speed is divided by four.\nThere is no hardware support for float64, it is done in software and greatly affects the execution speed.\n...\nThe SM needs to be fed instructions and data which resides in the GPU memory.\nTo avoid stalling, GPUs don't try to avoid memory trips with a lot of cache and speculation like CPUs do.\n...\nThe execution model still revolves around warps of 32 threads scheduled on a SM.\nOnly thanks to a process of 40nm, NVidia doubled/quadrupled everything.\nA SM can now schedule two half-warp (16 threads) simultaneously thanks to two arrays of 16 CUDA cores.\nWith each core executing one instruction per clock, a SM can retire one warp instruction per clock (4x the capacity of Tesla SM).\n...\nThere is a semi-hardware support for float64 where operations are carried by two CUDA core combined.\n...\nWith four warp scheduler able to process a whole warp in one clock (compared to Fermi's half-warp design) the SMX now contains 192 cores.\n...\nWith the release of Turing in 2018, Nvidia operated its \"biggest architectural leap forward in over a decade\"^[13]^.\nNot only the \"Turing SM\" added A.I dedicated Tensor cores, they also gained Raytracing cores.\n...\nBesides the new cores, Turing added three major features.\nFirst, the CUDA core is now a super-scalar able to execute both integer instruction and float instruction in parallel.\n...\nSecond, the new GDDR6X memory sub-system, backed by 16 controllers, can now achieve 14 Gbps.\nLast, threads are no longer sharing their Instruction Pointer in a warp.\nThanks to Independent Thread Scheduling introduced in Volta each thread has its own IP.\nAs a result, SMs are free to fine schedule threads in a warp without the need to make them converge as soon as possible.\n...\nThe next architecture, codenamed Ampere, is rumored to be announced later in 2020.",
"title": "A history of NVidia Stream Multiprocessor - Fabien Sanglard",
"url": "https://fabiensanglard.net/cuda/",
"date": "2020-05-02",
"last_updated": "2026-05-18",
"source": "web"
},
{
"id": 8,
"snippet": "According to Michael Houston from NVIDIA, Tensor Cores are specialized hardware units designed to accelerate mixed precision training.",
"title": "Tensor Cores Explained in Simple Terms - DigitalOcean",
"url": "https://www.digitalocean.com/community/tutorials/understanding-tensor-cores",
"date": "2025-08-04",
"last_updated": "2026-05-18",
"source": "web"
},
{
"id": 9,
"snippet": "2. You’ve already figured out the constant cache is 8kB per SM.\nIt’s not configurable (not sure what you would configure about it, anyway).",
"title": "Multiprocessor architecture - CUDA - NVIDIA Developer Forums",
"url": "https://forums.developer.nvidia.com/t/multiprocessor-architecture/159951",
"date": "2020-11-19",
"last_updated": "2026-05-19",
"source": "web"
},
{
"id": 10,
"snippet": "CUDA Cores are the fundamental processing units inside NVIDIA GPUs, and they handle parallel computations.\n...\nCUDA Cores are the generalists.\nThey handle standard floating-point operations, integer math, and general-purpose parallel computing tasks.\n...\nCUDA Cores handle general parallel computing tasks like data preprocessing and activation functions, while **Tensor Cores** specialize in the matrix multiplications that dominate neural network training.\nBoth work together to accelerate your complete machine learning pipeline.",
"title": "NVIDIA CUDA Cores: How They Work and Why They Matter (2026)",
"url": "https://www.thundercompute.com/blog/cuda-cores-explained-ai-training",
"date": "2026-05-17",
"last_updated": "2026-05-14",
"source": "web"
},
{
"id": 11,
"snippet": "The GPU’s memory bandwidth determines how fast it can move data from/to memory (vRAM) to the computation cores.\nIt’s a more representative indicator than GPU Memory Speed.\nIt is determined by the data transmission speed between memory and computation cores and the number of individual parallel links in the bus between these two parts.\n...\nGPU memory bandwidth refers to the rate at which data can be transferred between the GPU and its memory (VRAM).\nIt is measured in gigabytes per second (GB/s) and plays a critical role in handling large datasets, real-time rendering, and AI/ML workloads.",
"title": "GPU Memory Bandwidth and Its Impact on Performance - DigitalOcean",
"url": "https://www.digitalocean.com/community/tutorials/gpu-memory-bandwidth",
"date": "2025-08-05",
"last_updated": "2026-05-18",
"source": "web"
}
],
"type": "search_results",
"queries": [
"NVIDIA streaming multiprocessor architecture overview",
"NVIDIA CUDA cores and tensor cores definition",
"NVIDIA HBM memory bandwidth explanation"
]
},
{
"id": "msg_64a3d245-e4a5-494f-85ea-53211b74bf68",
"content": [
{
"text": "NVIDIA’s GPU compute model is built around many simple parallel cores grouped into streaming multiprocessors (SMs), backed by very high memory bandwidth (often HBM) so those cores stay busy. [web:1][web:11] SMs contain CUDA cores for general arithmetic and Tensor Cores for dense matrix math, all fed by wide, fast VRAM links. [web:2][web:5][web:6][web:10]\n\n## Streaming multiprocessors (SMs)\n\nAn SM is the fundamental **compute** block of an NVIDIA GPU, roughly analogous to a CPU core but designed to run thousands of lightweight threads in parallel. [web:1][web:5][web:7] Each SM has its own register file, caches/shared memory, warp schedulers, and execution units (CUDA cores, Tensor Cores, special‑function units, load/store units). [web:1][web:5][web:9]\n\nThe programming model launches many threads that are grouped into warps of 32 threads, and warps are scheduled onto SMs. [web:1][web:7] An SM can keep many warps resident (for example up to 64 warps / 2048 threads on recent parts) and quickly switch between them every cycle to hide memory latency instead of relying on deep caches or speculation. [web:1][web:5][web:7]\n\n## CUDA cores\n\n“CUDA cores” are the scalar FP32 execution units inside an SM that perform general-purpose parallel arithmetic. [web:4][web:10] In common usage, the advertised “N CUDA cores” number means N single‑precision units, each capable of roughly one fused multiply‑add (FMA) per cycle for floating‑point math. [web:4][web:2]\n\nCUDA cores handle the bulk of standard GPU work: floating‑point math, integer math, and general GPGPU tasks like elementwise ops, reductions, and control-heavy kernels. [web:4][web:10] An SM in a modern Ada architecture GPU, for example, might have 128 CUDA cores, and a large GPU aggregates many SMs to reach tens of thousands of CUDA cores in total. [web:2][web:7][web:10]\n\n## Tensor Cores\n\nTensor Cores are specialized matrix‑math units inside SMs designed to accelerate small matrix multiply‑accumulate operations, especially for deep learning. [web:6][web:8] They operate on tiles (e.g., 4×4, 16×16 depending on architecture) in mixed or reduced precision (FP16, BF16, TF32, INT8, etc.), delivering far higher throughput than doing the same operations on CUDA cores. [web:6][web:8]\n\nIn NVIDIA’s architectures since Volta/Turing, each SM includes multiple Tensor Cores alongside its CUDA cores (e.g., one Ada SM has 4 Tensor Cores and 128 CUDA cores). [web:2][web:7] The CUDA programming model or frameworks like cuDNN, TensorRT, and PyTorch map large GEMM/convolution ops into many tile operations, which Tensor Cores execute to accelerate training and inference. [web:6][web:8][web:10]\n\n## HBM and memory bandwidth\n\nGPU memory bandwidth is the rate at which data can be moved between GPU cores (inside SMs) and VRAM, typically measured in GB/s. [web:11] High bandwidth is crucial because massively parallel SMs can consume data extremely quickly and will stall if memory cannot supply operands fast enough. [web:1][web:11]\n\nHigh Bandwidth Memory (HBM) is a 3D‑stacked DRAM technology with a very wide bus that provides much higher bandwidth at lower power than traditional GDDR. [web:11] NVIDIA’s data‑center GPUs pair many SMs and Tensor Cores with HBM so that large models and tensors can be streamed efficiently, ensuring the arithmetic units (CUDA and Tensor Cores) can run near peak utilization rather than waiting on data. [web:1][web:6][web:11]",
"type": "output_text",
"annotations": [],
"logprobs": []
}
],
"role": "assistant",
"status": "completed",
"type": "message"
}
],
"status": "completed",
"error": null,
"usage": {
"input_tokens": 6685,
"output_tokens": 915,
"total_tokens": 7600,
"cost": {
"currency": "USD",
"input_cost": 0.00388,
"output_cost": 0.00915,
"total_cost": 0.01848,
"cache_creation_cost": null,
"cache_read_cost": 0.00045,
"tool_calls_cost": 0.005
},
"input_tokens_details": {
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 3584,
"cached_tokens": 3584
},
"tool_calls_details": {
"search_web": {
"invocation": 1
}
},
"output_tokens_details": {
"reasoning_tokens": 0
}
},
"background": false,
"completed_at": 1779391739,
"frequency_penalty": 0,
"incomplete_details": null,
"instructions": "## Abstract\n<role>\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n</role>\n\n## Instructions\n<tools_workflow>\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the <tool_call_limit> below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n<tool_call_limit> Make at most three tool calls before concluding.</tool_call_limit>\n</tools_workflow>\n\n## Citation Instructions\n<citation_instructions>\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n<common_source_types> are included below.\n\n<common_source_types>\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n</common_source_types>\n\n<formatting_citations>\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n</formatting_citations>\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n<response_guidelines>\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n<tone>\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n</tone>\n\n### Lists and Paragraphs\n<lists_and_paragraphs>\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n</lists_and_paragraphs>\n\n### Summaries and Conclusions\n<summaries_and_conclusions>\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n</summaries_and_conclusions>\n\n## Prohibited Meta-Commentary\n<prohibited_commentary>\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n</prohibited_commentary>\n\n<copyright_requirements>\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n</copyright_requirements>\n\nCurrent date: Thursday, May 21, 2026\n\n",
"max_output_tokens": 8192,
"max_tool_calls": null,
"metadata": {},
"parallel_tool_calls": true,
"presence_penalty": 0,
"previous_response_id": null,
"prompt_cache_key": null,
"reasoning": null,
"safety_identifier": null,
"service_tier": "default",
"store": true,
"temperature": 1,
"text": {
"format": {
"type": "text"
}
},
"tool_choice": "auto",
"tools": [
{
"type": "web_search"
},
{
"type": "fetch_url"
}
],
"top_logprobs": 0,
"top_p": 1,
"truncation": "disabled",
"user": null
}
Next Steps
Web Search
Use web search for source-grounded, current context.
Agent API Models
Browse available models and pricing across all supported providers.
Presets
Explore pre-configured setups for common use cases like low and medium.
Output Control
Configure streaming responses and structured outputs with JSON schema.
Model Fallback
Specify multiple models for automatic failover and higher availability.
Prompt Guide
Best practices for effective prompting with web search models.
Search Filters
Control search results with domain, date, and location filters.
API Reference
View complete endpoint documentation and parameters.
Need help? Check out our community for support and discussions with other developers.