On March 16, 2026, Mistral AI released Mistral Small 4, an open-weight model under the Apache 2.0 license that combines instruction following, step-by-step reasoning and coding in a single multimodal checkpoint. It uses a mixture-of-experts (MoE) design with 119B total parameters, of which only 6.5B are active per token, and it supports a 256K-token context window. For teams that previously ran separate chat, reasoning and coding models, Mistral Small 4 offers one permissively licensed model to deploy, monitor and fine-tune.
Key Facts#
- Release: March 16, 2026, announced in Mistral's platform changelog, with weights published on Hugging Face as
Mistral-Small-4-2603. - Architecture: a 119B-parameter MoE model with 6.5B active parameters per token.
- Capabilities: Mistral describes it as a "hybrid model unifying instruct, reasoning, and coding" that accepts text and image input.
- Context window: 256K tokens.
- License: Apache 2.0, with no usage restrictions beyond the standard license terms.
- API: available as
mistral-small-2603and the aliasmistral-small-latest, listed at $0.15 per million input tokens and $0.60 per million output tokens. - Hardware: Mistral's documentation lists minimum GPU memory of about 238 GB in BF16, 119 GB in FP8 and 60 GB in FP4.
What Happened#
Mistral announced the model with a one-line changelog entry: it had released Mistral Small 4, "a hybrid model unifying instruct, reasoning, and coding in a single multimodal model with a 256k context window." The same entry introduced Leanstral, which Mistral calls its first open-source code agent for formal proof engineering in the Lean 4 language.
Mistral's model documentation fills in the details. The API model supports chat completions, function calling, agents and conversations, connectors, structured outputs, predicted outputs, document question answering, prefix completion and batching. Input can be text or images, and output is text. Mistral's documentation for its Vibe coding CLI now recommends two open models for offline use: the dense Devstral Small 2 for lighter hardware, and Mistral Small 4 for hybrid reasoning and coding with long context.
The version number reflects a real change in design. Earlier Small releases were dense 24B models. Mistral Small 3.2, released in June 2025, had 24B parameters and a 128K context window. Moving to a sparse 119B MoE model lets Mistral raise capacity substantially while keeping the number of active parameters, and so the per-token compute, far below that of a dense model of similar size.
Background#
Mistral, the French AI lab, has maintained an unusually permissive open-weight line. During 2025 it shipped specialized 24B Apache 2.0 models for different jobs: Devstral Small for agentic software engineering, Magistral Small for multimodal reasoning, and Mistral Small 3.x for general chat. On December 2, 2025, it released Mistral Large 3, an open-weight multimodal MoE model with 675B total and 41B active parameters, together with the Ministral 3 family of 3B, 8B and 14B models, all under Apache 2.0.
Mistral Small 4 applies the Large 3 approach, a granular MoE with a small active footprint, to the company's mid-sized tier and folds the specialized models' skills into one checkpoint. That matches a broader 2026 trend. Hugging Face's summer 2026 report found that most large Chinese open releases in 2026 shipped under Apache 2.0 or MIT, and that developers increasingly standardize on model families that cover many sizes and tasks.
Why It Matters for Developers#
One model can now replace three. If your system routes requests between a chat model, a reasoning model and a coding model, each with its own deployment, prompts and evaluation suite, Mistral Small 4 lets you collapse that into one endpoint. That reduces GPU memory, deployment work and the number of model versions you have to track. In .NET, this fits naturally behind a single IChatClient from Microsoft.Extensions.AI, with middleware for logging, caching and function invocation.
Size it by total parameters. With 6.5B active parameters, Mistral Small 4 generates tokens quickly, but all 119B parameters must be resident. Mistral's figures suggest that an FP4 build fits on a single 80 GB GPU, while FP8 needs about 119 GB and therefore more than one card or a larger accelerator. Long contexts add KV-cache memory on top. It is a workstation or server model, not a laptop model. For smaller machines, the Ministral 3 models or the dense Devstral Small 2 are more realistic, as covered in our guide to local AI in .NET.
The Apache 2.0 license matters for European and regulated customers. It has no user thresholds, attribution rules or regional carve-outs, and the weights can run entirely inside your own infrastructure. Combined with a European vendor, that makes Mistral Small 4 a practical option when data residency or supplier sovereignty is a requirement.
The API pricing is low enough to use as a baseline. At $0.15 per million input tokens, the hosted model is cheap to benchmark against larger proprietary models. Track per-feature token costs with the techniques in our guide to LLM observability and cost control before you decide whether self-hosting pays off.
As with any hybrid model, test both modes. Reasoning improves hard tasks but increases latency and output tokens, and coding-agent behavior depends heavily on tool-calling reliability. Build a small evaluation set for each workload before you consolidate routes onto one model, following the patterns in our AI agent architecture guide.
What's Next#
Mistral's model documentation does not link a launch paper or benchmark report for Small 4, so independent benchmarks and community fine-tunes will shape how it is perceived. Community quantizations appeared on Hugging Face in several formats, including MLX builds for Apple silicon and GGUF files for llama.cpp-based runtimes.
Mistral has also signaled bigger open releases. In July 2026, Tech Times reported that a new Mistral open-weight model was entering early access with selected partners. Details such as parameter count, benchmarks and license terms had not been published at the time. For developers, the safe assumption is that Mistral will keep shipping Apache 2.0 weights across sizes, which makes it worth keeping a Mistral model in your evaluation matrix.