Using DeepSeek-V3.2 on Chutes

DeepSeek-V3.2 is DeepSeek's MIT-licensed flagship reasoning and agent model, built on a Mixture-of-Experts transformer with DeepSeek Sparse Attention for efficient long-context inference. On Chutes it is served inside a Trusted Execution Environment (TEE) through the standard OpenAI-compatible gateway, so any OpenAI SDK works with a one-line base URL change.

Overview

Released in December 2025, DeepSeek-V3.2 is a finetune of DeepSeek-V3.2-Exp-Base. The published FP8 safetensors weights total about 685B parameters across 61 layers, with 256 routed experts plus one shared expert and 8 experts activated per token. The context window is 163,840 tokens.

Its headline architectural feature is DeepSeek Sparse Attention (DSA): each query attends to only the top 2,048 most relevant prior tokens (index_topk in the config) instead of the full history, which sharply reduces attention cost in long-context scenarios while preserving quality. Post-training combines a scaled reinforcement-learning framework with a large agentic task-synthesis pipeline that generates tool-use training data at scale. DeepSeek reports performance comparable to GPT-5, and gold-medal results at the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI); the winning submissions are published in the repository for independent verification.

The chat template is new in this generation: it revises the tool-calling format, adds a "thinking with tools" capability that interleaves reasoning with tool calls, and introduces a developer role reserved exclusively for search-agent scenarios.

Model specifications

Quick start

Authenticate with Authorization: Bearer $CHUTES_API_KEY. The model name is deepseek-ai/DeepSeek-V3.2-TEE on the shared gateway https://llm.chutes.ai/v1.

curl -X POST "https://llm.chutes.ai/v1/chat/completions" \
  -H "Authorization: Bearer $CHUTES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "deepseek-ai/DeepSeek-V3.2-TEE",
  "messages": [{"role": "user", "content": "Prove that sqrt(2) is irrational."}],
  "stream": true,
  "max_tokens": 1024,
  "temperature": 1.0,
  "top_p": 0.95
}'

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://llm.chutes.ai/v1",
    api_key=os.environ["CHUTES_API_KEY"],
)

response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V3.2-TEE",
    messages=[{"role": "user", "content": "Prove that sqrt(2) is irrational."}],
    max_tokens=1024,
    temperature=1.0,
    top_p=0.95,
)
print(response.choices[0].message.content)

const res = await fetch("https://llm.chutes.ai/v1/chat/completions", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.CHUTES_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: "deepseek-ai/DeepSeek-V3.2-TEE",
    messages: [{ role: "user", content: "Prove that sqrt(2) is irrational." }],
    max_tokens: 1024,
    temperature: 1.0,
    top_p: 0.95,
  }),
});
const data = await res.json();
console.log(data.choices[0].message.content);

Parameters and tuning

Chute defaults (from the live endpoint definition): temperature 0.7, max_tokens 1024, seed 42, stream true on the streaming variant. The /v1/completions endpoint additionally accepts the full vLLM sampling surface: top_p (default 1), top_k (default -1, disabled), min_p, presence_penalty, frequency_penalty, repetition_penalty, logprobs, and more.

DeepSeek's model card recommends temperature = 1.0, top_p = 0.95, which also matches the repository's generation_config.json. Because the chute default temperature is 0.7, pass these explicitly when you want DeepSeek's reference behavior. Lower temperatures tighten output for deterministic extraction tasks; the recommended settings suit reasoning and agentic work, where the RL post-training was tuned. Raise max_tokens well above the 1024 default for reasoning-heavy prompts, since thinking content consumes output budget.

What it's best at

DeepSeek-V3.2 is built for reasoning-heavy and agentic workloads:

• Agentic tool-use loops. The thinking-with-tools chat template and large-scale agentic post-training make it a strong engine for multi-step tool orchestration, and the developer role supports search-agent pipelines specifically.
• Long-context work. Sparse attention keeps inference efficient at 100K+ token inputs, useful for repository-scale code analysis or long-document reasoning within the 163,840-token window.
• Competition-grade math and code. The model card reports IMO 2025 and IOI 2025 gold-medal performance and GPT-5-comparable reasoning.
• Complex multi-step problem solving where an RL-trained reasoner beats plain instruct models.

It is not the right pick for vision or audio input (text-only), for latency-critical short completions where a small dense model is cheaper, or when you specifically want the high-compute Speciale variant, which is reasoning-only, lacks tool calling, and is a separate deployment.

How Chutes serves this model

This chute runs the model inside a Trusted Execution Environment: inference executes on attested confidential-compute hardware, so prompts and outputs are processed inside the enclave. TEE serving is a deployment property and does not change model behavior. The upstream weights are already FP8 (e4m3 block-quantized) as published by DeepSeek.

Serving is vLLM-based on the shared OpenAI-compatible gateway, exposing /v1/chat/completions and /v1/completions (both with streaming) plus GET /v1/models. Billing follows Chutes' standard per-token LLM pricing. See the model page, the machine-readable llms.txt, and the callable OpenAPI spec.

FAQ

What context window does DeepSeek-V3.2 support?

163,840 tokens (max_position_embeddings in the upstream config). DeepSeek Sparse Attention is specifically designed to keep long-context inference efficient, so the model remains practical near the top of that window.

Does DeepSeek-V3.2 support function calling and tool use?

Yes. This generation introduces a revised tool-calling format and a thinking-with-tools mode in its chat template, and the model was post-trained on large-scale synthesized agentic tasks. Send OpenAI-style tools in your chat completion request via the gateway.

Can I use DeepSeek-V3.2 commercially?

Yes. Both the repository and the model weights are released under the MIT license, which permits commercial use, modification, and redistribution.

What does the TEE suffix mean?

The chute runs inference inside a Trusted Execution Environment, i.e. attested confidential-compute hardware. Prompts and outputs are processed inside the enclave. It is a serving-level property and does not change the model's behavior or output quality.

How do I call it from the OpenAI SDK?

Point the client at base_url https://llm.chutes.ai/v1 with your Chutes API key, and set model to deepseek-ai/DeepSeek-V3.2-TEE. Both chat completions and legacy completions work, and streaming is supported.

What sampling settings should I use?

DeepSeek recommends temperature 1.0 and top_p 0.95, which also match the repo's generation_config. The chute default is temperature 0.7, so pass the recommended values explicitly if you want reference behavior.

Is DeepSeek-V3.2 a reasoning model?

Yes, it is post-trained with reinforcement learning for reasoning and emits thinking content via its chat template. DeepSeek reports performance comparable to GPT-5, with the separate high-compute Speciale variant going further on pure reasoning (but without tool calling).

Sources

• deepseek-ai/DeepSeek-V3.2 on Hugging Face
• Upstream config.json
• Upstream generation_config.json
• DeepSeek-V3.2-TEE model page on Chutes
• Live llms.txt for this chute