Cerebras API

Cerebras provides the world's fastest AI inference (2,000+ tokens/s). OpenAI-compatible API with Python and TypeScript SDKs.

Quick Reference

| Resource | Location | |----------|----------| | API Base URL | https://api.cerebras.ai/v1 | | Get API Key | https://cloud.cerebras.ai | | Python SDK | pip install cerebras_cloud_sdk | | TypeScript SDK | npm install @cerebras/cerebras_cloud_sdk |

Available Models

Deprecation Notice: qwen-3-32b and llama-3.3-70b are scheduled for deprecation on February 16, 2026.

Production Models

Fully supported for production use.

| Model | Model ID | Parameters | Speed | |-------|----------|------------|-------| | Llama 3.1 8B | llama3.1-8b | 8B | ~2200 tok/s | | Llama 3.3 70B | llama-3.3-70b | 70B | ~2100 tok/s | | OpenAI GPT OSS | gpt-oss-120b | 120B | ~3000 tok/s | | Qwen 3 32B | qwen-3-32b | 32B | ~2600 tok/s |

Preview Models

For evaluation only - may be discontinued with short notice.

| Model | Model ID | Parameters | Speed | |-------|----------|------------|-------| | Qwen 3 235B Instruct | qwen-3-235b-a22b-instruct-2507 | 235B | ~1400 tok/s | | Z.ai GLM 4.7 | zai-glm-4.7 | 355B | ~1000 tok/s |

Migrating to GLM? See GLM 4.7 Migration Guide.

Model Selection Guide

| Use Case | Recommended Model | |----------|-------------------| | Speed-critical (real-time chat) | llama3.1-8b | | Balanced (chat, coding, math) | llama-3.3-70b | | Hybrid reasoning | qwen-3-32b | | Multilingual, instruction following | qwen-3-235b-a22b-instruct-2507 | | Science, math, complex reasoning | gpt-oss-120b | | Agents, superior tool use | zai-glm-4.7 |

Model Compression

All models are unpruned original versions. Precision varies:

| Model | Precision | Weights | |-------|-----------|---------| | llama3.1-8b | FP16 | HuggingFace | | llama-3.3-70b | FP16 | HuggingFace | | gpt-oss-120b | FP16/FP8 (weights only) | HuggingFace | | qwen-3-32b | FP16 | HuggingFace | | qwen-3-235b-a22b-instruct-2507 | FP16/FP8 (weights only) | HuggingFace | | zai-glm-4.7 | FP16/FP8 (weights only) | HuggingFace |

Note: FP16/FP8 models use selective weight-only quantization for storage. Sensitive layers remain at full precision, with dequantization on-the-fly. Activations and KV cache remain unquantized.

Basic Usage

Python

import os
from cerebras.cloud.sdk import Cerebras

client = Cerebras(api_key=os.environ.get("CEREBRAS_API_KEY"))

response = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=[{"role": "user", "content": "Explain quantum computing"}]
)
print(response.choices[0].message.content)

TypeScript

import Cerebras from '@cerebras/cerebras_cloud_sdk';

const client = new Cerebras({ apiKey: process.env.CEREBRAS_API_KEY });

const response = await client.chat.completions.create({
    model: 'llama-3.3-70b',
    messages: [{ role: 'user', content: 'Explain quantum computing' }]
});
console.log(response.choices[0].message.content);

Streaming

The Cerebras API supports streaming responses, allowing messages to be sent back in chunks and displayed incrementally as they are generated. Set stream=True to receive an iterable of chunks.

Python

stream = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=[{"role": "user", "content": "Why is fast inference important?"}],
    stream=True
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

TypeScript

const stream = await client.chat.completions.create({
    model: 'llama-3.3-70b',
    messages: [{ role: 'user', content: 'Why is fast inference important?' }],
    stream: true
});

for await (const chunk of stream) {
    process.stdout.write(chunk.choices[0]?.delta?.content || '');
}

Streaming Notes

• Each chunk contains a delta object with incremental content
• usage and time_info are only available in the final chunk
• Use flush=True in Python print for real-time display: print(..., end="", flush=True)

Cancel Streaming (TypeScript)

const stream = await client.chat.completions.create({
    model: 'llama-3.3-70b',
    messages: [{ role: 'user', content: 'Long response' }],
    stream: true
});

for await (const chunk of stream) {
    if (shouldStop) {
        stream.controller.abort();
        break;
    }
    process.stdout.write(chunk.choices[0]?.delta?.content || '');
}

Async Streaming (Python)

from cerebras.cloud.sdk import AsyncCerebras

client = AsyncCerebras()

async def stream_response():
    stream = await client.chat.completions.create(
        model="llama-3.3-70b",
        messages=[{"role": "user", "content": "Tell me a joke"}],
        stream=True
    )

    async for chunk in stream:
        content = chunk.choices[0].delta.content
        if content:
            print(content, end="", flush=True)

Tool Calling

Tool calling (also known as function calling) enables models to interact with external tools, APIs, or applications to perform actions and access real-time information.

Supported models: gpt-oss-120b, qwen-3-32b, zai-glm-4.7

How It Works

1. Define tools - Provide name, description, and parameters for each tool
2. Send request - Include tool definitions with your API call
3. Model decides - Model analyzes if a tool can help answer the question
4. Execute & respond - Your code executes the tool and returns results to the model

Define Tools

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "strict": True,
            "description": "Get temperature for a given location.",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "City and country e.g. Toronto, Canada"
                    }
                },
                "required": ["location"],
                "additionalProperties": False
            }
        }
    }
]

Make API Call

response = client.chat.completions.create(
    model="zai-glm-4.7",
    messages=[{"role": "user", "content": "What's the weather in Tokyo?"}],
    tools=tools
)

Handle Tool Calls

import json

choice = response.choices[0].message

if choice.tool_calls:
    # Add assistant message with tool_calls
    messages.append(choice)

    for tool_call in choice.tool_calls:
        # Execute your tool
        arguments = json.loads(tool_call.function.arguments)
        result = get_weather(arguments["location"])

        # Append tool result
        messages.append({
            "role": "tool",
            "content": json.dumps(result),
            "tool_call_id": tool_call.id
        })

    # Get final response
    final_response = client.chat.completions.create(
        model="zai-glm-4.7",
        messages=messages
    )
    print(final_response.choices[0].message.content)

Parallel Tool Calling

When a query requires multiple independent data points (e.g., comparing weather in different cities), the model can request multiple tools at once.

response = client.chat.completions.create(
    model="zai-glm-4.7",
    messages=[{"role": "user", "content": "Is Toronto warmer than Montreal?"}],
    tools=tools,
    parallel_tool_calls=True  # Default: enabled
)

# Response may contain multiple tool_calls
for tool_call in response.choices[0].message.tool_calls:
    print(f"Tool: {tool_call.function.name}, Args: {tool_call.function.arguments}")

To disable parallel calling:

response = client.chat.completions.create(
    model="zai-glm-4.7",
    messages=messages,
    tools=tools,
    parallel_tool_calls=False  # Force sequential execution
)

TypeScript Example

const tools: Cerebras.Chat.ChatCompletionTool[] = [{
    type: 'function',
    function: {
        name: 'get_weather',
        strict: true,
        description: 'Get temperature for a given location.',
        parameters: {
            type: 'object',
            properties: {
                location: { type: 'string', description: 'City name' }
            },
            required: ['location'],
            additionalProperties: false
        }
    }
}];

const response = await client.chat.completions.create({
    model: 'zai-glm-4.7',
    messages: [{ role: 'user', content: 'Weather in Paris?' }],
    tools
});

if (response.choices[0].message.tool_calls) {
    for (const toolCall of response.choices[0].message.tool_calls) {
        const args = JSON.parse(toolCall.function.arguments);
        // Execute tool and continue conversation
    }
}

Best Practices

• Use strict: true for reliable JSON argument parsing
• Always set additionalProperties: false in parameter schemas
• Provide clear, descriptive tool descriptions
• Handle cases where model doesn't call any tools

Structured Outputs

Generate structured data with enforced JSON schema compliance. Key benefits:

• Reduced Variability - Consistent outputs adhering to predefined fields
• Type Safety - Enforces correct data types, preventing mismatches
• Easier Parsing - Direct use in applications without extra processing

Defining the Schema

Define a JSON schema specifying fields, types, and required properties.

For every required array, you must set additionalProperties: false.

Python (with Pydantic)

from pydantic import BaseModel

class Movie(BaseModel):
    title: str
    director: str
    year: int

response = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=[
        {"role": "system", "content": "You are a helpful assistant"},
        {"role": "user", "content": "Suggest a sci-fi movie"}
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "movie",
            "strict": True,
            "schema": Movie.model_json_schema()
        }
    }
)

import json
movie = json.loads(response.choices[0].message.content)

Python (with raw schema)

movie_schema = {
    "type": "object",
    "properties": {
        "title": {"type": "string"},
        "director": {"type": "string"},
        "year": {"type": "integer"}
    },
    "required": ["title", "director", "year"],
    "additionalProperties": False
}

response = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=[{"role": "user", "content": "Suggest a sci-fi movie"}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "movie",
            "strict": True,
            "schema": movie_schema
        }
    }
)

TypeScript (with Zod)

import { z } from 'zod';
import { zodToJsonSchema } from 'zod-to-json-schema';

const MovieSchema = z.object({
    title: z.string(),
    director: z.string(),
    year: z.number().int()
});

const response = await client.chat.completions.create({
    model: 'llama-3.3-70b',
    messages: [{ role: 'user', content: 'Suggest a sci-fi movie' }],
    response_format: {
        type: 'json_schema',
        json_schema: {
            name: 'movie',
            strict: true,
            schema: zodToJsonSchema(MovieSchema)
        }
    }
});

const movie = MovieSchema.parse(JSON.parse(response.choices[0].message.content || '{}'));

Response Format Modes

| Mode | Valid JSON | Adheres to Schema | Extra Fields | Constrained Decoding | |------|------------|-------------------|--------------|----------------------| | json_schema (strict: true) | Yes | Yes (guaranteed) | No | Yes | | json_schema (strict: false) | Yes (best-effort) | Yes | Yes | No | | json_object | Yes | No (flexible) | No | No |

Enabling each mode:

• Strict: response_format: { type: "json_schema", json_schema: { strict: true, schema: ... } }
• Non-strict: response_format: { type: "json_schema", json_schema: { strict: false, schema: ... } }
• JSON object: response_format: { type: "json_object" }

Schema Requirements

• Set "additionalProperties": false for all objects with required fields
• Max nesting: 5 levels
• Max schema length: 5,000 chars
• No recursive schemas

tools and response_format cannot be used in the same request.

Reasoning Models

Reasoning models generate intermediate thinking tokens before their final response, enabling better problem-solving and allowing inspection of the model's thought process.

Supported models: qwen-3-32b, gpt-oss-120b, zai-glm-4.7

Reasoning Format Options

| Format | Behavior | Use Case | |--------|----------|----------| | parsed | Reasoning in separate reasoning field, logprobs split into reasoning_logprobs | When you need structured access to thinking | | raw | Reasoning prepended to content with wrapper tokens (<think>...</think> for GLM/Qwen) | When you want full visibility | | hidden | Reasoning text dropped from response (tokens still counted/billed) | When you want benefits without exposing thinking | | none | Uses model's default behavior | Default |

Default behaviors by model:

• Qwen3: raw (or hidden for JSON output)
• GLM: text_parsed
• GPT-OSS: text_parsed

Basic Usage

response = client.chat.completions.create(
    model="qwen-3-32b",
    messages=[{"role": "user", "content": "Solve: 15% of 240"}],
    reasoning_format="parsed"
)
print("Thinking:", response.choices[0].message.reasoning)
print("Answer:", response.choices[0].message.content)

const response = await client.chat.completions.create({
    model: 'qwen-3-32b',
    messages: [{ role: 'user', content: 'Solve: 15% of 240' }],
    reasoning_format: 'parsed'
});
console.log('Thinking:', response.choices[0].message.reasoning);
console.log('Answer:', response.choices[0].message.content);

GPT-OSS: Reasoning Effort

Control reasoning intensity with reasoning_effort:

response = client.chat.completions.create(
    model="gpt-oss-120b",
    messages=[{"role": "user", "content": "Prove the Pythagorean theorem"}],
    reasoning_effort="high"  # "low", "medium" (default), "high"
)

GLM: Disable Reasoning

Toggle reasoning on/off for GLM:

response = client.chat.completions.create(
    model="zai-glm-4.7",
    messages=[{"role": "user", "content": "Quick factual question"}],
    disable_reasoning=True  # Skip thinking for simple queries
)

Multi-Turn Reasoning Context

To retain reasoning awareness across conversation turns, include prior reasoning in assistant messages using the model's native format.

GPT-OSS (reasoning prepended directly):

messages = [
    {"role": "user", "content": "What is 25 * 4?"},
    {"role": "assistant", "content": "Multiply 25 times 4 equals 100. The answer is 100."},
    {"role": "user", "content": "Now divide that by 2."}
]
response = client.chat.completions.create(model="gpt-oss-120b", messages=messages)

GLM/Qwen (reasoning in <think> tags):

messages = [
    {"role": "user", "content": "What is 25 * 4?"},
    {"role": "assistant", "content": "<think>Multiply 25 times 4 equals 100.</think>The answer is 100."},
    {"role": "user", "content": "Now divide that by 2."}
]
response = client.chat.completions.create(model="zai-glm-4.7", messages=messages)

Predicted Outputs

Reduce latency by specifying parts of the response that are already known. (Public Preview)

Supported models: gpt-oss-120b, llama3.1-8b, zai-glm-4.7

Predicted Outputs speed up response generation when parts of the output are already known. This is most useful when regenerating text or code that requires only minor changes.

Python

code = """
html {
    margin: 0;
    padding: 0;
    box-sizing: border-box;
    color: #00FF00;
}
"""

response = client.chat.completions.create(
    model="gpt-oss-120b",
    messages=[
        {"role": "user", "content": "Change the color to blue. Respond only with code."},
        {"role": "user", "content": code}
    ],
    prediction={"type": "content", "content": code}
)

TypeScript

const code = `
html {
    margin: 0;
    padding: 0;
    box-sizing: border-box;
    color: #00FF00;
}
`;

const response = await client.chat.completions.create({
    model: 'gpt-oss-120b',
    messages: [
        { role: 'user', content: "Change the color to blue. Respond only with code." },
        { role: 'user', content: code }
    ],
    prediction: { type: 'content', content: code }
});

Token-Reuse Metrics

The response includes usage metrics showing prediction efficiency:

{
  "usage": {
    "completion_tokens": 224,
    "prompt_tokens": 204,
    "completion_tokens_details": {
      "accepted_prediction_tokens": 76,
      "rejected_prediction_tokens": 20
    }
  }
}

A high ratio of accepted to rejected tokens indicates efficient prediction reuse.

Best Practices

• Use when most output is known - The larger the known section, the greater the efficiency gain
• Set temperature=0 - Reduces randomness and increases token acceptance
• Keep predictions accurate - Misaligned predictions increase rejected tokens
• Monitor metrics - Track accepted vs rejected tokens to evaluate effectiveness

Limitations

• Rejected tokens are billed at completion-token rates
• Not compatible with: logprobs, n > 1, tools
• Reasoning tokens may generate additional rejected_prediction_tokens

Prompt Caching

Store and reuse previously processed prompts to reduce latency. Designed to significantly reduce Time to First Token (TTFT) for long-context workloads like multi-turn conversations, RAG, and agentic workflows.

How It Works

Automatic - No code changes required. Works on all supported API requests.

1. Prefix Matching - System analyzes the beginning of your prompt (system prompts, tool definitions, few-shot examples)
2. Block-Based Caching - Prompts processed in blocks (100-600 tokens). Matching blocks reuse cached computation
3. Cache Hit - Cached blocks skip processing, resulting in lower latency
4. Cache Miss - Prompt processed normally, prefix stored for future matches
5. Auto Expiration - TTL guaranteed 5 minutes, may persist up to 1 hour

The entire beginning of your prompt must match exactly with a cached prefix. Even a single character difference causes a cache miss.

Checking Cache Usage

Check the usage.prompt_tokens_details.cached_tokens field in your response:

response = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=messages
)

cached = response.usage.prompt_tokens_details.cached_tokens
print(f"Cached tokens: {cached}")

{
  "usage": {
    "prompt_tokens": 1500,
    "prompt_tokens_details": {
      "cached_tokens": 1200
    }
  }
}

Best Practices for Cache Hits

• Keep prefixes consistent - System prompts, tool definitions, and few-shot examples should be identical across requests
• Order matters - Place stable content (system prompt, tools) before dynamic content (user messages)
• Multi-turn conversations - Cache naturally builds as conversation history grows
• RAG workflows - Place frequently-used context at the beginning

Example: Multi-Turn with Tools

# System message and tools are cached across turns
messages = [
    {"role": "system", "content": "You are a shopping assistant."},
    {"role": "user", "content": "Where is my order ORD-123456?"}
]

# Turn 1 - creates cache for system + tools
response = client.chat.completions.create(
    model="qwen-3-32b",
    messages=messages,
    tools=tools
)
print(f"Turn 1 cached: {response.usage.prompt_tokens_details.cached_tokens}")

# Turn 2 - reuses cached system + tools
messages.append(response.choices[0].message)
messages.append({"role": "user", "content": "Please cancel it."})

response = client.chat.completions.create(
    model="qwen-3-32b",
    messages=messages,
    tools=tools
)
print(f"Turn 2 cached: {response.usage.prompt_tokens_details.cached_tokens}")

FAQ

• Pricing: No additional cost. Standard token rates apply
• Quality: Caching only affects input processing. Output generation unchanged
• Manual clear: Not available. System manages cache automatically
• TTL: Guaranteed 5 minutes, up to 1 hour depending on load

Async Usage (Python)

import asyncio
from cerebras.cloud.sdk import AsyncCerebras

client = AsyncCerebras()

async def main():
    response = await client.chat.completions.create(
        model="llama-3.3-70b",
        messages=[{"role": "user", "content": "Hello"}]
    )
    print(response.choices[0].message.content)

asyncio.run(main())

Error Handling

All errors inherit from cerebras.cloud.sdk.APIError. Main categories:

• APIConnectionError - Unable to connect to the API
• APIStatusError - API returns non-success status code (4xx or 5xx)

Error Codes

| Status | Exception | Description | |--------|-----------|-------------| | 400 | BadRequestError | Invalid request parameters | | 401 | AuthenticationError | Invalid or missing API key | | 402 | PaymentRequired | Payment required | | 403 | PermissionDeniedError | Insufficient permissions | | 404 | NotFoundError | Resource not found | | 422 | UnprocessableEntityError | Validation error | | 429 | RateLimitError | Too many requests | | 500 | InternalServerError | Server error | | 503 | ServiceUnavailable | Service temporarily unavailable | | N/A | APIConnectionError | Network/connection issue |

Python Example

import cerebras.cloud.sdk
from cerebras.cloud.sdk import Cerebras

client = Cerebras()

try:
    response = client.chat.completions.create(
        model="llama-3.3-70b",
        messages=[{"role": "user", "content": "Hello"}]
    )
except cerebras.cloud.sdk.APIConnectionError as e:
    print("Server could not be reached")
    print(e.__cause__)
except cerebras.cloud.sdk.RateLimitError as e:
    print("Rate limited - implement backoff")
except cerebras.cloud.sdk.APIStatusError as e:
    print(f"Error {e.status_code}: {e.response}")

TypeScript Example

try {
    const response = await client.chat.completions.create({
        model: 'llama-3.3-70b',
        messages: [{ role: 'user', content: 'Hello' }]
    });
} catch (err) {
    if (err instanceof Cerebras.APIError) {
        console.log(err.status);   // 400
        console.log(err.name);     // BadRequestError
        console.log(err.headers);  // Response headers
    } else {
        throw err;
    }
}

Retries & Timeouts

Automatic Retries

By default, these errors are retried 2 times with exponential backoff:

• Connection errors
• 408 Request Timeout
• 429 Rate Limit

• = 500 Internal errors

# Python - configure retries
client = Cerebras(max_retries=0)  # Disable retries

# Per-request override
client.with_options(max_retries=5).chat.completions.create(...)

// TypeScript - configure retries
const client = new Cerebras({ maxRetries: 0 });

// Per-request override
await client.chat.completions.create(params, { maxRetries: 5 });

Timeouts

Default timeout is 60 seconds. On timeout, APITimeoutError is thrown.

# Python - configure timeout
client = Cerebras(timeout=20.0)  # 20 seconds

# Granular control
import httpx
client = Cerebras(
    timeout=httpx.Timeout(60.0, read=5.0, write=10.0, connect=2.0)
)

# Per-request override
client.with_options(timeout=5.0).chat.completions.create(...)

// TypeScript - configure timeout
const client = new Cerebras({ timeout: 20 * 1000 });

// Per-request override
await client.chat.completions.create(params, { timeout: 5 * 1000 });

TCP Warming

SDK sends warmup requests on init to reduce first-token latency. Disable if needed:

client = Cerebras(warm_tcp_connection=False)

const client = new Cerebras({ warmTCPConnection: false });

Key Parameters

| Parameter | Description | |-----------|-------------| | model | Model identifier (required) | | messages | Conversation history (required) | | temperature | Randomness 0-1.5 (default varies) | | max_completion_tokens | Max output tokens | | stop | Up to 4 stop sequences | | stream | Enable streaming | | response_format | text, json_object, or json_schema | | tools | Function definitions for tool calling | | reasoning_format | parsed, raw, hidden | | reasoning_effort | low, medium, high (gpt-oss only) |

References

For detailed SDK documentation:

• Python: See references/python.md
• TypeScript: See references/typescript.md