Mistral Small 3.2 24B Instruct
Mistral's open-weight 24B instruction-tuned model with tool use, web search, and structured output — a production-ready upgrade to Small 3.1.
Overview
What is Mistral Small 3.2 24B Instruct
Mistral Small 3.2 24B Instruct is a 24-billion-parameter language model released by Mistral AI in June 2025. It is an open-weight, Apache 2.0-licensed update to Small 3.1 with improved instruction following, reduced repetition, stronger function calling, and support for tool use and structured output.
Running it privately on Venice
On Venice, Mistral Small 3.2 runs with zero retention — your prompts are not stored, profiled, or used for training. It supports tool use, web search, and structured JSON output, giving you open-weight sovereignty and permissionless inference without Big-Tech surveillance. Per-token billing means you pay only for what you generate, with no subscription lock-in.
Assessment
Strengths and limitations
- Open weights under Apache 2.0: self-host, fine-tune, or redistribute without licensing friction.
- Major instruction-following upgrade over 3.1: Arena Hard v2 rose to 43.1% and Wildbench v2 to 65.33%.
- Production reliability: infinite-generation rate dropped to 1.29%, cutting repetition loops roughly in half.
- Strong coding performance for its size: HumanEval Plus 92.90% and MBPP Plus 78.33%.
- Native tool use and web search on Venice, plus structured JSON output for agent pipelines.
- Extremely low inference cost on Venice: among the cheapest capable open-weight models available.
- General knowledge is solid but not frontier-level — MMLU sits at ~80.5% and MMLU-Pro at ~69%, behind larger flagship models.
- 24B parameters means it can trail larger closed models on the hardest STEM and multi-step agentic tasks.
- Not uncensored: it retains safety alignment and will refuse certain harmful requests.
- It has no knowledge of events after its training data; use the built-in web search on Venice for real-time facts.
- Runs quantized to fp8 on Venice, which trades a small amount of precision for speed and cost.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- Mistral AI
- Released
- June 20, 2025
- Architecture
- Dense transformer
- Parameters
- 24B
- License
- Apache 2.0
- Open weights
- Yes — downloadable from Hugging Face
- Context window
- 256K tokens
- Max output
- 16.384K tokens
- Capabilities
- Function calling, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Jan 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mistral-small-3-2-24b-instruct",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.09 per 1M input tokens and $0.25 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Context window | Best for | Open weights | Price (Venice) |
|---|---|---|---|---|
| Mistral Small 3.2 24B Instruct | 256K tokens | Tool use & cost-efficient agents | Yes | $0.09 in · $0.25 out / 1M |
| DeepSeek V3.2 | 160K tokens | Reasoning & STEM | Yes | $0.33 in · $0.48 out / 1M |
| Google Gemma 4 31B Instruct | 256K tokens | General instruction & fine-tuning | Yes | $0.12 in · $0.36 out / 1M |
| Kimi K2.6 | 256K tokens | Long-document analysis | Yes | $0.75 in · $3.50 out / 1M |
The best-value open-weight model for tool use and structured output under $0.30/1M tokens.
Use cases
What it is good for
- 01Cost-efficient agents and tool-calling pipelines that need structured JSON output.
- 02High-volume text classification, summarization, and data extraction.
- 03Coding assistance and lightweight software engineering (HumanEval+ 92.90%).
- 04Multilingual chat and content generation across 20+ languages.
- 05Low-cost inference workloads where open-weight sovereignty is required.
Prompting
Getting better results
Be explicit and precise in prompts — the 3.2 update is specifically tuned for exact instruction following.
Use structured output schemas for reliable JSON extraction; the model supports schema constraints on Venice.
For coding tasks, provide clear function signatures and test cases to leverage its strong pass rates.
Enable web search when you need facts after the model's training cutoff instead of relying on parametric knowledge.
If you hit repetition, raise the temperature slightly or add a stop sequence — the model already reduces loops by default.
Version history
Predecessor with the same base architecture.
Current — improved instruction following, reduced repetition, and stronger function calling.
FAQ
Frequently asked questions
It is a 24-billion-parameter dense language model released by Mistral AI in June 2025. It is an open-weight, Apache 2.0-licensed update to Small 3.1 with improved instruction following, reduced repetition, stronger function calling, and support for tool use and structured output.
Venice bills per token at $0.09 per 1M input tokens and $0.25 per 1M output tokens. There is no subscription required, and you can start with free credits.
Yes — the weights are released under Apache 2.0 on Hugging Face, so you can self-host, fine-tune, and redistribute them. Inference on Venice is private and permissionless.
Yes. On Venice it supports function calling, web search, and structured JSON output via schema constraints, making it suitable for agent pipelines.
Choose Mistral Small 3.2 for low-cost tool use, structured output, and high-volume agents. Choose DeepSeek V3.2 if you need stronger deep reasoning and coding performance and can pay roughly 3× the per-token cost.
No. It is open-weight but not uncensored — it retains safety alignment and will refuse certain harmful requests. Venice does not add extra content filters beyond the model's own weights.
It supports more than 20 languages including English, French, German, Spanish, Portuguese, Italian, Japanese, Korean, Russian, Chinese, Arabic, Hindi, Bengali, and others.
It runs under Venice's private, zero-retention tier — prompts are not stored, profiled, or used for training. Venice does not run this model in a TEE or end-to-end encrypted session, but your data is not retained.
Run Mistral Small 3.2 24B Instruct privately
No prompt logging. No data used for training.