Mistral Small 4 is an open-weight, multimodal Mixture-of-Experts language model from Mistral AI, released on 16 March 2026 under the Apache 2.0 licence, that combines instruct, reasoning and coding in one model. This profile collects the published facts about it in one place: API identifier, context window, parameter counts, prices, reasoning modes, hardware guidance, where it can be run, provider-reported benchmarks, data handling and deprecation status. It is written for developers, architects, buyers and governance teams who want checked numbers rather than launch coverage. Every fact comes from Mistral AI's own documentation, model card, launch post, Hugging Face model card, pricing page or help centre, listed under Standards and references. Where Mistral AI publishes nothing on a point, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Mistral AI's public documentation as of November 2026. Model details change, so confirm against Mistral AI's current documentation.
Key takeaways
- Released 16 March 2026. The Mistral changelog and launch post record Mistral Small 4 (API ID mistral-small-2603, version 26.03) on that date, and the model card lists it as GA.
- 256k context, 119B total and 6.5B active parameters. The model card gives 119B total with 6.5B active; the Hugging Face card describes 128 experts with 4 active per token, and text and image input with text output.
- $0.15 input and $0.60 output per million tokens. The launch post labels these figures; the model card lists the same two numbers without labels. Batch processing is 50% off according to the pricing page.
- Apache 2.0 open weights. Weights are on Hugging Face as mistralai/Mistral-Small-4-119B-2603 in BF16 and FP8, with an NVFP4 checkpoint and an EAGLE speculative decoding head.
- One model replaces several. Mistral's models overview names Small 4 as the replacement for Mistral Small 3.0 to 3.2, Magistral Small 1.0 to 1.2 and Devstral Small 1.1; Small 4 itself has no deprecation date.
What is Mistral Small 4?
Mistral Small 4 is the current model in Mistral AI's Small line. The model card describes it as "Our powerful hybrid model unifying instruct, reasoning, and coding capabilities in a single model." The launch post says it is the first Mistral model to combine the capabilities of Magistral (reasoning), Pixtral (multimodal) and Devstral (agentic coding), so that "one model now delivers all three."
It accepts text and images and returns text. A per-request reasoning_effort parameter switches between fast replies in the style of Mistral Small 3.2 (none) and step-by-step reasoning similar in verbosity to earlier Magistral models (high). Because it is a Mixture-of-Experts model, only a fraction of its parameters are active per token; see Mixture of Experts Explained. For the full Mistral line-up, see Mistral AI Models.
Key facts at a glance
As of 2026-11-15. Sources: Mistral Small 4 model card (docs.mistral.ai), Hugging Face model card, Mistral changelog and launch post.
| Field | Value | Source |
|---|---|---|
| Provider | Mistral AI | Model card |
| API ID | mistral-small-2603 (the card shows one further alias without naming it) | Model card; launch post |
| Version | 26.03 | Model card; models overview |
| Release date | 2026-03-16 | Model card; changelog; launch post |
| Status | GA, current | Model card; models overview |
| Context window | 256k tokens | Model card; changelog; launch post |
| Max output | Not stated by Mistral AI on the pages checked (2026-11-15) | Model card |
| Knowledge cutoff | Not stated by Mistral AI on the pages checked (2026-11-15) | Model card; Hugging Face card |
| Input modalities | Text, image | Launch post; Hugging Face card |
| Output modalities | Text | Launch post; Hugging Face card |
| Reasoning effort | none, high | Launch post; Hugging Face card |
| Open weights | Yes, mistralai/Mistral-Small-4-119B-2603 (BF16 and FP8) | Hugging Face card |
| Licence | Apache 2.0 | Model card; launch post |
| Parameters | 119B total, 6.5B active (model card); the launch post gives 6B active, 8B including embedding and output layers | Model card; launch post |
| Languages | Dozens, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic | Hugging Face card |
Release and version history
As of 2026-11-15. Dates from the Mistral changelog, models overview and model card.
| Date | Event | Source |
|---|---|---|
| 2024-02-26 | mistral-small-latest first introduced, then pointing to mistral-small-2402 (Small 1.0) | Changelog |
| 2025-11-30 | Mistral Small 3.0 and 3.1, and Magistral Small 1.0 and 1.1, retired, with Small 4 named as replacement | Models overview |
| 2026-03-16 | Mistral Small 4 (mistral-small-2603) released, multimodal with a 256k context window | Changelog; launch post |
| 2026-04-30 | Mistral Small 3.2 and Magistral Small 1.2 deprecated, replacement Small 4 | Models overview |
| 2026-05-31 | Devstral Small 1.1 retired, replacement Small 4 | Models overview |
| 2026-07-31 | Mistral Small 3.2 and Magistral Small 1.2 retired | Models overview |
Mistral publishes one version of Small 4, 26.03. The replacement mappings show how Mistral folded its separate reasoning and small coding lines into this single model. See LLM Naming and Versioning Explained for how the year.month version labels work.
How much does Mistral Small 4 cost?
As of 2026-11-15. USD per million tokens. Sources: Mistral Small 4 model card, launch post and mistral.ai/pricing.
| Item | Price | Notes | Source |
|---|---|---|---|
| Input | $0.15 | Labelled as input in the launch post; the model card lists $0.15 without a label and offers a EUR view | Launch post; model card |
| Output | $0.60 | Labelled as output in the launch post; reasoning tokens are generated as output | Launch post; model card |
| Cached input | Not stated by Mistral AI on the pages checked (2026-11-15) | Pricing page says cached input reduces input cost by up to 90% in general | mistral.ai/pricing |
| Batch | 50% reduction | General statement on the pricing page, not model specific | mistral.ai/pricing |
| Self-hosted weights | No per-token fee from Mistral; own compute cost | Apache 2.0 permits commercial use | Hugging Face card; launch post |
Worked example at list rates: a request with 100,000 input tokens and 10,000 output tokens costs $0.015 for input plus $0.006 for output, $0.021 in total. With reasoning_effort set to high, output includes the reasoning trace, so measure tokens per task in both modes. See Model Pricing Explained and How to set spend limits on LLM API keys.
Context window and output limits
The context window is 256k tokens on the model card, the changelog and the launch post. Mistral's published vLLM example sets max-model-len to 262,144 tokens, which matches 256k. Mistral does not publish a separate maximum output figure on the pages checked.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window | 256k tokens | 262,144 in the vLLM example | Model card; Hugging Face card |
| Max output | Not stated by Mistral AI on the pages checked (2026-11-15) | None | Model card |
| API rate limits | Not stated by Mistral AI on the pages checked (2026-11-15) | Set per workspace in Mistral Studio | Model card |
| Minimum self-hosting hardware | 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200 | As stated by Mistral | Launch post |
| Recommended self-hosting hardware | 4x HGX H100, 4x HGX H200 or 2x DGX B200 | As stated by Mistral | Launch post |
Modalities and tools
Small 4 takes text and images and returns text. The launch post tags it as text-to-text, agentic, multimodal and lightweight.
- API features listed on the model card: chat completions, function calling, agents and conversations, built-in tools, structured outputs, predicted outputs, document Q&A, prefix and batching (/v1/batch).
- Reasoning effort: none (fast replies, matching the chat style of Mistral Small 3.2) or high (step-by-step reasoning). Mistral recommends high for complex prompts.
- Recommended sampling: temperature 0.7 with reasoning_effort high; between 0.0 and 0.7 with none, depending on the task (Hugging Face card).
- Tool use: native function calling and JSON output.
- Self-hosting stacks named by Mistral: vLLM (recommended), SGLang, llama.cpp, LM Studio and Transformers; fine-tuning through Axolotl; an EAGLE head for speculative decoding and an NVFP4 checkpoint for 4-bit serving.
- Efficiency claims: Mistral reports a 40% reduction in end-to-end completion time in a latency-optimised setup and 3x more requests per second in a throughput-optimised setup, both versus Mistral Small 3.
Built-in tools and agents give the model reach into other systems. Treat each tool as a credentialed integration with a least-privilege scope. See the Open-Weight Model Security Guide for self-hosted deployments.
Where is Mistral Small 4 available?
As of 2026-11-15. Channels as named in the launch post, the model card and the Hugging Face card.
| Channel | Details | Source |
|---|---|---|
| Mistral API and AI Studio | mistral-small-2603 | Model card; launch post |
| Hugging Face weights | mistralai/Mistral-Small-4-119B-2603, plus NVFP4 and EAGLE repositories | Hugging Face card |
| NVIDIA | Free prototyping on build.nvidia.com, day-0 NVIDIA NIM for production, NVIDIA NeMo for customisation | Launch post |
| Le Chat | Not stated by Mistral AI on the pages checked (2026-11-15) | Launch post |
| Amazon Bedrock, Google Vertex AI, Microsoft Foundry | Not stated by Mistral AI on the pages checked (2026-11-15) | Launch post |
Access restrictions
Small 4 is not a restricted-access model. It is GA on the Mistral API and its weights are public on Hugging Face under Apache 2.0, which permits commercial and non-commercial use. The Hugging Face card adds that the model must not be used to infringe third-party rights. Mistral states no Model-as-a-Service condition for Small 4 on the pages checked.
Benchmarks (provider-reported, dated)
As of 2026-11-15. All figures are as reported by Mistral AI in the Small 4 launch post (16 March 2026) or Hugging Face model card. Several results appear only as charts without numbers in the text; those are described, not quoted.
| Evaluation | Mistral Small 4 | Comparator in source | Source and date |
|---|---|---|---|
| GPQA Diamond | 71.2 | None given in text | Hugging Face card |
| AA LCR (long-context reasoning) | 0.72, with about 1.6K characters of output | Mistral states Qwen models need about 5.8K to 6.1K characters for comparable performance | Launch post; Hugging Face card |
| LiveCodeBench | Outperforms GPT-OSS 120B (no number in text) | About 20% less output than GPT-OSS 120B | Launch post; Hugging Face card |
| End-to-end completion time | 40% lower | Versus Mistral Small 3, latency-optimised setup | Launch post |
| Requests per second | 3x higher | Versus Mistral Small 3, throughput-optimised setup | Launch post |
The launch post states that with reasoning enabled Small 4 matches or surpasses GPT-OSS 120B on the three benchmarks it charts; AIME25 appears only as an image. Other comparators named are Mistral Small 3.2 24B Instruct and Magistral Small 2509. Benchmark settings differ between labs, so treat cross-lab comparisons with care.
Model card and system card
Mistral publishes two model cards for Small 4: the documentation card at docs.mistral.ai, with the API name, version, status, pricing, features and licence, and the Hugging Face card, with architecture, recommended settings and deployment guidance. Mistral does not publish a separate system card or safety evaluation for Small 4 on the pages checked. What the cards state:
- Design: a hybrid of instruct, reasoning and coding, replacing separate Magistral and Devstral Small models.
- Architecture: Mixture-of-Experts with 128 experts, 4 active per token; tagged mistral3, served in Transformers as Mistral3ForConditionalGeneration.
- Weights formats: Safetensors in BF16 and FP8 (F8_E4M3), plus a separate NVFP4 checkpoint.
- Usage condition: Apache 2.0, with a statement that the model must not be used to infringe third-party rights.
- Parameter wording: the documentation card says 6.5B active; the launch post says 6B active, or 8B including embedding and output layers.
For what model and system cards usually contain, see Model Cards and System Cards Explained.
Data handling
Mistral AI's help centre applies the same rules to Small 4 as to other API models; there is no Small 4 specific data policy on the pages checked.
- Free mode in Mistral Studio: Mistral states it may use inputs and outputs to train its models, with an opt-out available.
- Pay-as-you-go: Mistral states customers keep control over this processing and can opt out at any time.
- Zero Data Retention: available only on pay-as-you-go and only for stateless API calls (chat completions, FIM, embeddings, moderation and classification, OCR, audio). It is granted on request, at Mistral's discretion.
- ZDR exclusions: agents, batch files, conversations, libraries, the Files API, Vibe Work and Chat. Small 4's agents, conversations, built-in tools and batching features therefore sit outside ZDR.
- Self-hosted weights: data stays wherever you run the model; retention is your own responsibility.
Mistral API keys are non-human identities with direct spend and data exposure. See Zero Data Retention for AI and the LLMjacking Guide.
Deprecation status
As of 15 November 2026, Mistral's models overview lists Mistral Small 4 as a current model with no deprecation or retirement date. It is the named replacement for several retired small, reasoning and coding models. Watch the models overview and changelog for changes.
| Model or item | Details | Source |
|---|---|---|
| Mistral Small 4 (mistral-small-2603) | Current; no deprecation or retirement date listed (2026-11-15) | Models overview |
| Mistral Small 3.2 (mistral-small-2506) | Deprecated 2026-04-30, retired 2026-07-31; replacement Small 4 | Models overview |
| Mistral Small 3.0 and 3.1 | Deprecated 2025-11-06, retired 2025-11-30; replacement Small 4 | Models overview |
| Magistral Small 1.0, 1.1 and 1.2 | Retired 2025-11-30 (1.0, 1.1) and 2026-07-31 (1.2); replacement Small 4 | Models overview |
| Devstral Small 1.1 (devstral-small-2507) | Deprecated 2026-02-27, retired 2026-05-31; replacement Small 4 | Models overview |
How Mistral Small 4 differs from other models in its family
As of 2026-11-15. Sources: Mistral model cards, models overview and changelog. Models listed by release date.
| Model | Released | Context | Parameters | Licence | Status |
|---|---|---|---|---|---|
| Mistral Small 3.2 | Version 25.06 | Not stated on the pages checked | 24B (as named in the Small 4 Hugging Face card) | Not stated on the pages checked | Retired 2026-07-31 |
| Mistral Small 4 | 2026-03-16 | 256k | 119B total, 6.5B active | Apache 2.0 | GA |
| Mistral Medium 3.5 | Version 26.04 | Not stated on the pages checked | Not stated by Mistral AI | Not stated on the pages checked | Deprecated, retires 2027-04-30 |
| Mistral Large 4 | Preview 2026-10-06; GA 2026-11-05 | 1M | 1.05T total, 52B active | Mistral Ecosystem Open-Weight License | GA |
The documented differences from Large 4 are context (256k versus 1M), scale, reasoning modes (Small 4 offers none and high; Large 4 offers low, medium and high), licence (Apache 2.0 versus the Mistral Ecosystem Open-Weight License) and list price ($0.15 and $0.60 versus $0.70 and $2.10). Predicted outputs and built-in tools are listed on the Small 4 card. See Mistral Large 4, Reasoning Models Explained and Open-Weight Frontier Models Compared.
Practitioner checklist
- Call mistral-small-2603 explicitly in production code and record the date you adopted it.
- If you still reference Mistral Small 3.x, Magistral Small or Devstral Small 1.1, migrate to Small 4 and retest prompts; those models are retired.
- Set reasoning_effort per workload (none or high) and measure output tokens and latency in both modes.
- Size self-hosted clusters against Mistral's stated minimum and recommended hardware before committing.
- Check whether your use relies on agents, conversations, batch or built-in tools, which sit outside Zero Data Retention.
- Scope function-calling and built-in tools to least privilege, with approval for destructive actions.
- Store Mistral API keys in a secrets manager, separate workspaces by environment and set spend limits.
- Watch the models overview and changelog for a deprecation notice on Small 4.
Standards and references
- Mistral AI: Mistral Small 4 model card (accessed 2026-11-15)
- Mistral AI on Hugging Face: Mistral-Small-4-119B-2603 model card (accessed 2026-11-15)
- Mistral AI: Mistral Small 4 launch post (accessed 2026-11-15)
- Mistral AI: Changelog (accessed 2026-11-15)
- Mistral AI: Models overview (accessed 2026-11-15)
- Mistral AI: Pricing (accessed 2026-11-15)
- Mistral AI Help Center: Can I activate Zero Data Retention (ZDR)? (accessed 2026-11-15)
- Mistral AI Help Center: Do you use my user data to train your models? (accessed 2026-11-15)
Related NHI Mgmt Group resources: Mistral AI Models, Mistral Large 4, Open-Weight Frontier Models Compared, Open-Weight Model Licence Tracker, LLM Pricing Tracker.