Mistral Small 4
Mistral Small 4 unifies instruct, reasoning, and vision in a single open, efficient MoE model — deployable on-premise or via API with configurable reasoning effort.
Overview
What is Mistral Small 4
Mistral Small 4 is a 119B-parameter Mixture-of-Experts (MoE) model from Mistral AI, released in March 2026. It combines instruction, reasoning, and multimodal vision capabilities into one unified model, with only 6.5B active parameters per token for high efficiency. Released under Apache 2.0, it supports open deployment and configurable reasoning modes.
Running it privately on Venice
On Venice, Mistral Small 4 runs with zero retention — your prompts are never stored or used for training. This open, uncensored model can be used privately for agentic coding, document analysis, or vision tasks without surveillance. You maintain sovereignty over your data while accessing a frontier-class MoE architecture.
Assessment
Strengths and limitations
- Unified model for instruct, reasoning, and vision — eliminates need for multiple specialized models.
- Efficient MoE architecture activates only ~6.5B of 119B parameters per token, reducing compute cost.
- Configurable reasoning effort: toggle between fast responses and deep reasoning per request.
- Open weights under Apache 2.0: fully auditable, self-hostable, and customizable.
- Supports vision, function calling, web search, and code-optimized workflows natively.
- Local deployment still resource-intensive despite sparse activation.
- Documentation and tooling assume advanced technical proficiency.
- No native no-code interface: best suited for developers and ML teams.
- Higher API cost than some lightweight open models on sustained high-volume workloads.
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
- March 16, 2026
- Architecture
- Mixture of Experts (128 experts, 4 active per token)
- Parameters
- 119B total, ~6.5B active per token
- Open weights
- Yes — Apache 2.0
- Context window
- 256K tokens
- Max output
- 65.536K tokens
- Capabilities
- Vision, Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Mar 2026
- License
- Apache 2.0
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-2603",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.19 per 1M input tokens and $0.75 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 | Max resolution | Strongest at | Open weights | Price (Venice) |
|---|---|---|---|---|
| Mistral Small 4 | N/A | Efficiency & multimodal reasoning | Yes | $0.19 in · $0.75 out / 1M |
| DeepSeek V4 Flash 0731 | N/A | Speed & low-cost throughput | No | $0.17 in · $0.35 out / 1M |
| Google Gemma 4 31B Instruct | N/A | On-device & edge deployment | Yes | $0.12 in · $0.36 out / 1M |
| Claude Sonnet 4.6 | N/A | Balanced performance & cost | No | $3.60 in · $18 out / 1M |
Open, efficient MoE model with configurable reasoning and vision.
Use cases
What it is good for
- 01Autonomous coding agents that require both reasoning and code generation.
- 02Multimodal document analysis combining text and image inputs.
- 03Private AI assistants where data sovereignty and zero retention are required.
- 04On-premise deployment for regulated or European data environments.
- 05Agentic workflows with dynamic reasoning depth based on task complexity.
Prompting
Getting better results
Use the `reasoning_effort` parameter to control depth — 'low' for chat, 'high' for math or planning.
Interleave images and text in your input for multimodal analysis.
Leverage function calling for tool integration — works seamlessly in reasoning mode.
For cost efficiency, use speculative decoding with the Mistral eagle head checkpoint.
Version history
Predecessor — dense 24B model without multimodal or reasoning modes.
Current — unified MoE model with vision, reasoning, and open weights.
FAQ
Frequently asked questions
Mistral Small 4 is a 119B-parameter Mixture-of-Experts model from Mistral AI, released in March 2026. It unifies instruction, reasoning, and vision capabilities into one open, efficient model with configurable reasoning depth and multimodal input support.
On Venice, Mistral Small 4 costs $0.19 per 1M input tokens and $0.75 per 1M output tokens. There is no subscription — you pay only for the tokens you use.
Mistral Small 4 is open source under the Apache 2.0 license — you can inspect, modify, and self-host it. However, API usage on Venice is billed per token, not free.
Yes, Mistral Small 4 supports vision — it accepts both text and image inputs, enabling multimodal analysis, document understanding, and visual reasoning tasks.
Yes, Mistral Small 4 supports function calling and tool use, allowing it to act as an agentic coding assistant or integrate with external APIs and systems.
Mistral Small 4 is better for reasoning, vision, and open deployment; DeepSeek V4 Flash is faster and cheaper per token but closed and less capable in complex reasoning tasks.
Yes — Mistral Small 4 is open weights under Apache 2.0, so you can deploy it on-premise or in private cloud environments. Quantized versions are available for efficiency.
Mistral Small 4 has a 256K token context window, supporting long-form conversations, document analysis, and complex agentic workflows.
Run Mistral Small 4 privately
No prompt logging. No data used for training.