Mistral AI

Mistral Large 4

Mistral AI

Mistral Large 4, also called Le Chonk, is Mistral’s trillion-parameter multimodal flagship, first offered as a public API preview on October 6, 2026. It targets software engineering, long-running agents, visual grounding, and specialist enterprise work including finance, law, and cybersecurity. Training took place on Mistral’s European infrastructure. Its planned weight release is later in October; the preview available at the audit date does not include downloadable weights. Architecture and post-training details beyond the mixture-of-experts configuration remain limited.

Leanstral 1.5

Mistral AI

Leanstral 1.5 is Mistral’s open-weight specialist model for Lean 4 formal mathematics and program verification. The 1.5 update improves proof engineering through mid-training, supervised tuning, and reinforcement learning in theorem-proving and code-agent environments. Its first API release is dated June 30 in the changelog, before the July 2 announcement. It edits proof files, runs tools, and revises proofs after compiler feedback within an agent scaffold. Apache-2.0 weights permit local deployment and adaptation. Published success rates involve explicit inference budgets and multiple attempts; a long agent run is not the same as a single context window.

Mistral Medium 3.5

Mistral AI

Mistral Medium 3.5 is a dense open-weight multimodal model that combines instruction following, coding, and adjustable reasoning in a single checkpoint. It launched in public preview as the default model in Le Chat and Mistral Vibe, supporting long-running software and productivity tasks, parallel tool calls, and structured outputs. A newly trained vision encoder handles variable image dimensions and aspect ratios. The modified MIT license must be checked for commercial restrictions; the published four-GPU deployment is a configuration example rather than a universal memory requirement.

Leanstral

Mistral AI

Leanstral is Mistral’s open-weight specialist model for Lean 4 formal mathematics and program verification. The first Leanstral release focuses on realistic formal repositories and uses Lean compiler feedback as a correctness check. It edits proof files, runs tools, and revises proofs after compiler feedback within an agent scaffold. Apache-2.0 weights permit local deployment and adaptation. Published success rates involve explicit inference budgets and multiple attempts; a long agent run is not the same as a single context window.

Mistral Small 4

Mistral AI

Mistral Small 4 unifies chat instruction following, reasoning, coding, and image understanding in one configurable open-weight model. Its sparse expert architecture uses 128 experts, selecting four per token; the provider distinguishes six billion active expert parameters from eight billion including embedding and output layers. Users can choose direct answers or deeper reasoning per request. It is distributed under Apache 2.0 with support in common self-hosting runtimes. Despite the Small name, its total weights require considerably more memory than the earlier 24B Small models.

Mistral Small Creative

Mistral AI

Mistral Small Creative is an experimental Labs language model specializing in creative writing, narrative generation, roleplay, and character dialogue. It also supports general instruction following, structured output, document questions, and function calling through Mistral’s API. The public card documents a 32K context but does not publish a reliable parameter count or network architecture. This was a named public model release rather than a prompt template. The service was deprecated in March 2026; its historical publication remains part of the catalog. No downloadable weights were released.

Devstral 2

Mistral AI

Devstral 2 is Mistral’s second-generation open-weight coding model. The 123B dense model supports long-context repository exploration, multi-file editing, and agentic software development through Mistral Vibe or other coding scaffolds. Its modified MIT license contains commercial conditions. This page describes the large text model; the separately positioned 24B Devstral Small 2 has image input and an Apache-2.0 license. Published agent benchmark results rely on the specific scaffold and evaluation configuration.

Devstral Small 2

Mistral AI

Devstral Small 2 is the local deployment tier of Mistral’s second-generation coding family. The 24B dense model combines repository-level coding, tool interaction, and image input, allowing screenshots and visual requirements to inform software changes. Apache-2.0 weights enable private customization and self-hosting. It shares the large model’s 256K context capacity, but its smaller size trades difficult-task capability for lower memory and inference cost. Tool execution is provided by a coding scaffold such as Mistral Vibe.

Ministral 3

Mistral AI

Ministral 3 is Mistral’s third-generation family of small dense language and vision models, released alongside Mistral Large 3. Three sizes cover local, edge, and production deployments. Each size has base, instruct, and reasoning checkpoints with image understanding. The reasoning versions spend additional inference time on difficult tasks, while instruct models emphasize concise, inexpensive responses. All sizes are Apache-2.0 licensed; the 14B checkpoint documented here supports 256K context, with deployment memory varying by checkpoint and precision.

Mistral Large 3

Mistral AI

Mistral Large 3 is the open-weight flagship of the Mistral 3 release. Its sparse mixture-of-experts model combines multilingual text generation with image understanding and long-context instruction following. Mistral released base and instruction-tuned checkpoints under Apache 2.0, including compressed formats for more practical deployment. The release is an instruction model; a future reasoning edition mentioned in the announcement is not part of this record. Self-hosting still requires substantial memory and distributed serving.

Mistral Medium 3.1

Mistral AI

Mistral Medium 3.1 is a proprietary multimodal language model positioned for enterprise use. The separately released August update improves instruction following and enterprise assistant behavior over Medium 3. It accepts text and images and produces text, supporting document understanding and software development workflows. Hosted access is available through Mistral; enterprise deployment options should not be confused with freely downloadable weights. Public architectural and parameter details were not disclosed for this generation.

Voxtral

Mistral AI

Voxtral is Mistral’s open-weight audio-language family for understanding speech as well as transcribing it. The 3B Mini and 24B Small models answer questions, summarize recordings, and use tools directly from audio input, with multilingual support. Both are Apache-2.0 licensed and have hosted API access. The separately optimized Mini Transcribe service is a transcription-only product and is not a separate language model page here. Audio understanding generates text; this release is not a text-to-speech system.

Devstral Medium

Mistral AI

Devstral Medium is the larger hosted tier of Mistral’s first-generation software-engineering model family. Released alongside Devstral Small 1.1, it targets difficult repository tasks with multi-step tool interaction and code editing. It is accessible through the Mistral API, while its weights and architectural size were not publicly disclosed. The model’s performance reflects its surrounding coding agent and execution environment; it does not directly execute repository operations without supplied tools.

Mistral Small 3.2

Mistral AI

Mistral Small 3.2 is a permissively licensed small dense model for practical generative AI applications. Improves instruction following, function calling, and resistance to repetitive generations over the prior 3.1 checkpoint. It supports deployment on local hardware with suitable quantization and memory, as well as hosted inference. The model is useful for private assistants and domain adaptation, but the small parameter budget does not eliminate hallucinations or the need to evaluate specialized workloads. Base and instruct variants share this catalog identity where released.

Magistral Medium

Mistral AI

Magistral Medium is Mistral’s enterprise hosted reasoning tier, designed to produce explicit multi-step analysis across languages. It targets mathematics, business analysis, and coding tasks where extra inference time can improve answers. The Medium tier is proprietary; public weights and parameter counts were not disclosed. Later 1.1 and 1.2 updates are represented as release variants. Longer reasoning traces can add latency and remain fallible even when they sound systematic.

Magistral Small

Mistral AI

Magistral Small is Mistral’s open-weight reasoning tier, designed to produce explicit multi-step analysis across languages. It targets mathematics, business analysis, and coding tasks where extra inference time can improve answers. The 24B Small weights are Apache-2.0 licensed and permit local adaptation. Later 1.1 and 1.2 updates are represented as release variants. Longer reasoning traces can add latency and remain fallible even when they sound systematic.

Devstral Small

Mistral AI

Devstral Small is the original 24B agentic software-engineering model developed by Mistral with All Hands AI. It was trained to resolve real repository issues within scaffolds such as OpenHands, exploring codebases and making multi-file changes rather than only completing isolated functions. Apache-2.0 weights support private local deployment. The separately published July 10 Small 1.1 update remains represented as a variant of this initial small generation. Results depend on the agent scaffold, repository environment, and tools supplied to the model.

Mistral Medium 3

Mistral AI

Mistral Medium 3 is a proprietary multimodal language model positioned for enterprise use. The original Medium 3 release targets professional coding and visual understanding at lower cost than larger frontier models. It accepts text and images and produces text, supporting document understanding and software development workflows. Hosted access is available through Mistral; enterprise deployment options should not be confused with freely downloadable weights. Public architectural and parameter details were not disclosed for this generation.

Mistral Small 3.1

Mistral AI

Mistral Small 3.1 is a permissively licensed small dense model for practical generative AI applications. Adds image understanding, stronger text performance, and a longer 128K context to the Small series. It supports deployment on local hardware with suitable quantization and memory, as well as hosted inference. The model is useful for private assistants and domain adaptation, but the small parameter budget does not eliminate hallucinations or the need to evaluate specialized workloads. Base and instruct variants share this catalog identity where released.

Mistral Saba

Mistral AI

Mistral Saba is a regional language specialist trained on Middle Eastern and South Asian data. It emphasizes Arabic and Indian languages, including Tamil, and local cultural knowledge. Its 24B size enables lower-latency enterprise applications and customer-premises deployments under commercial terms. Public API access does not make the weights publicly downloadable. The model is intended for regional assistants and adaptations, and should be evaluated in the actual dialect and domain of each deployment.

Mistral Small 3

Mistral AI

Mistral Small 3 is a permissively licensed small dense model for practical generative AI applications. A latency-focused 24B release for the common instruction-following and conversational workload. It supports deployment on local hardware with suitable quantization and memory, as well as hosted inference. The model is useful for private assistants and domain adaptation, but the small parameter budget does not eliminate hallucinations or the need to evaluate specialized workloads. Base and instruct variants share this catalog identity where released.

Pixtral Large

Mistral AI

Pixtral Large is Mistral’s larger vision-language model, combining a 123B language decoder with a 1B vision encoder. It processes interleaved text and images for document interpretation, chart analysis, and visual reasoning, with a 128K context window. Downloadable weights are offered under the Mistral Research License, with separate commercial licensing required. The model is derived from Mistral Large 2 and complements the earlier 12B Pixtral. Visual grounding and extracted text can still be inaccurate on complex documents.

Mistral Large (123B)

Mistral AI

Flagship 123B-parameter model from Mistral AI, progressively upgraded through 2024 versions. Targeted to be a top-performing open model competing with Meta's Llama family on knowledge and reasoning tasks.

Pixtral 123B

Mistral AI

A 123B-parameter **multimodal** model from Mistral, combining language with vision ("Pix-"). A 12B smaller version was released under Apache 2.0. Pixtral extends Mistral's LLM capability to image understanding and description tasks.

Ministral 3B

Mistral AI

The smallest model in Mistral AI's Ministral family, part of 'les Ministraux'. A state-of-the-art edge model optimized for on-device computing with exceptional knowledge, commonsense, reasoning, and function-calling capabilities in the sub-10B category.

Ministral 8B

Mistral AI

A state-of-the-art 8B parameter model from Mistral AI's Ministral family, part of 'les Ministraux'. Features special interleaved sliding-window attention pattern for faster and memory-efficient inference, designed for edge computing and on-device applications.

Mistral Large 2

Mistral AI

Mistral Large 2 is the major 123B update to Mistral’s flagship language model, emphasizing code generation, mathematics, multilingual instruction following, and tool calling. It expanded context to 128K tokens and introduced downloadable weights under the Mistral Research License, with commercial use requiring a separate license. Later 24.11 checkpoint improvements remain part of this tier. Large 2 is a text model; the related Pixtral Large release supplies vision capabilities under its own identity.

Mistral NeMo

Mistral AI

Mistral NeMo is a 12B multilingual language model developed jointly by Mistral and NVIDIA. It provides base and instruction-tuned Apache-2.0 weights and long context for local and enterprise applications. The Tekken tokenizer improves token efficiency for many languages and code. It was designed for practical deployment with quantization support, but real memory use still depends on precision and context length. Its bilingual and multilingual quality should be measured on the actual languages required by an application.

Codestral Mamba

Mistral AI

Codestral Mamba is Mistral’s 7B code-generation model built on the Mamba state-space architecture. Its release explores efficient sequence modeling for long code inputs and fast local inference. Apache-2.0 weights and inference support allow adaptation without the original Codestral license restrictions. Mistral tested in-context retrieval up to 256K tokens, but that test should not be interpreted as guaranteed accuracy on every long code task. The model is a coding specialist rather than a general multimodal assistant.

Mathstral

Mistral AI

Mathstral is Mistral’s open-weight 7B mathematics and scientific reasoning specialist, released with the Project Numina collaboration. It builds on Mistral 7B and is intended for mathematical problem solving and adaptation in education and research. The published release supports 32K context and Apache-2.0 licensing. Multiple-sample voting and reward-based selection used in some reported evaluations are inference strategies, not intrinsic guarantees of a single generated solution.

Codestral

Mistral AI

Codestral is Mistral’s dedicated code-generation family, introduced as an open-weight 22B model trained across more than eighty programming languages. The original 24.05 checkpoint supports 32K context, completion, and fill-in-the-middle workflows under the Mistral AI Non-Production License. Later 25.01 and 25.08 service checkpoints improve low-latency code completion and fill-in-the-middle; their date-stamped refreshes remain variants of this named coding tier. Their deployment limits and weight availability differ from the original downloadable checkpoint. Codestral Mamba uses a distinct state-space architecture and has a separate page.

Mixtral 8×22B

Mistral AI

A sparse Mixture-of-Experts (SMoE) model that uses only 39B active parameters out of 141B total, offering unparalleled cost efficiency for its size. Sets a new standard for performance and efficiency within the AI community.

Mixtral 8×7B

Mistral AI

A sparse Mixture-of-Experts model (8 experts) with a total of 46.7B parameters (12.9B active per token). Outperformed dense models like Llama 2 70B and GPT-3.5 on many benchmarks at significantly lower compute cost.

Mistral 7B

Mistral AI

7.3B-parameter model trained from scratch with improved data and architecture enhancements. Open-source (Apache 2.0) with strong performance that rivaled larger 13B+ models on many benchmarks.