Mistral AI releases Mistral Small 4

What was released

On 16 March 2026, Mistral AI released Mistral Small 4 under the Apache 2.0 license ^[1]. The model is a 119B mixture-of-experts model with 6B active parameters and a 256K context window ^[1]. In a mixture-of-experts model, the total parameter count describes the full set of learned weights, while the active parameter count describes how much of the network is used for each token. That design lets a model carry a large total capacity without activating all of it on every step. A 256K context window means the model can take in a very long input in one request, such as a long document or a large codebase, subject to the serving system's limits.

Mistral describes Small 4 as combining reasoning, image input and agentic coding in one model ^[1]. Reasoning here means multi-step problem solving rather than a single short answer. Image input makes the model multimodal: it can receive images as part of a request, not only text. Agentic coding refers to coding workflows in which the model can act in steps, use tools and carry a task through an editing or debugging loop. Putting these abilities in one model means users do not have to choose a separate model for each of those three tasks, at least according to the release description ^[1].

Price and comparison

The API price is $0.15 per million input tokens and $0.60 per million output tokens ^[1]. Token pricing is quoted per million tokens, so the input price applies to text and other content sent to the model, and the output price applies to what the model generates. Mistral also compared Small 4 with GPT-OSS 120B. The company said Small 4 matches or surpasses GPT-OSS 120B on all three of its benchmarks while generating significantly shorter outputs ^[1]. That comparison is Mistral's own statement; it is not an independent evaluation.

Same-day announcements

The release did not stand alone. The same day, Mistral announced a partnership with NVIDIA to accelerate open frontier models and released Leanstral, a Lean 4 proof agent ^[2]. Lean 4 is a proof assistant and programming language used for formal mathematics and verification; a proof agent built for Lean 4 is aimed at constructing or checking formal proofs in that environment. Mistral also said it joined the NVIDIA Nemotron Coalition as a founding member ^[1]. The NVIDIA partnership, the Leanstral release and the coalition membership were announced alongside Small 4.

What the details show

The concrete details are the license, the architecture, the context length, the combined capabilities and the price. Apache 2.0 is a permissive open-source license that allows use, modification and distribution under its terms. A 119B total parameter count with 6B active parameters puts Small 4 in the mixture-of-experts category, where only part of the network runs for each token ^[1]. The 256K context and image input broaden the kinds of requests the model can accept ^[1]. The pricing gives a clear unit cost for API use ^[1]. The comparison with GPT-OSS 120B and the shorter outputs are part of Mistral's own message ^[1]. The same-day NVIDIA partnership, Leanstral release and Nemotron Coalition membership show that the announcement was part of a larger set of open-model activities ^[2][1].