#10Mistral AI

Mistral OCR 4

Mistral's fourth-generation document-understanding model. It turns PDF, Word, PowerPoint and OpenDocument files into structured output with bounding boxes, block labels and confidence scores, across 170 languages. It can be self-hosted in one container. A 4.1 update followed on 16 July 2026, and the API's latest alias now points to it.

• Parses PDF, DOC, PPT and OpenDocument files
• Bounding boxes, block classification, inline confidence scores
• 170 languages across 10 language groups
• Lab-stated OlmOCRBench 85.20, OmniDocBench 93.07

AccessRetired 30 Sep 2026 and replaced by Mistral OCR 4.1 (released 16 Jul 2026, generally available 31 Aug 2026) on the Mistral API and Studio. Was also on Amazon SageMaker and Microsoft Foundry. Not open weights.

Mistral Medium 3.5

Mistral's current flagship: a 128B dense multimodal model that merges instruction-following, reasoning and coding in one set of weights, with reasoning effort set per request. It is released as open weights under a modified MIT licence, replaces Mistral Medium 3.1 and Devstral 2, and powers Mistral's Vibe coding agents.

• 128B dense parameters, 256k-token context
• Text and image input via custom vision encoder
• Configurable reasoning effort per request
• Lab-stated 77.6% on SWE-Bench Verified
• Self-hostable on as few as four GPUs

AccessLe Chat and Mistral Vibe; Mistral API (mistral-medium-3-5-26-04, public preview at launch); open weights on Hugging Face under a modified MIT licence; NVIDIA build.nvidia.com and NIM.

Price$1.50 input / $7.50 output per 1M tokens

Voxtral TTS

Mistral's first text-to-speech model, with 4B parameters. It speaks nine languages and can adapt to a custom voice from about three seconds of reference audio. Weights are published for non-commercial use, and commercial use goes through Mistral's API. Mistral reports that human raters preferred its naturalness to ElevenLabs Flash v2.5.

• 4B parameters
• 9 languages incl. English, French, Hindi, Arabic
• Voice adaptation from ~3 s of reference audio
• ~70 ms latency for a 10 s sample, 500 characters
• Generates up to two minutes of audio natively

AccessMistral API and Studio playground; Le Chat; open weights on Hugging Face (mistralai/Voxtral-4B-TTS-2603) under CC BY-NC 4.0 (non-commercial).

Mistral Small 4

A 119B-parameter mixture-of-experts model with about 6B parameters active per token. It combines instruct, reasoning and coding in one model, accepts text and images, and has a 256k-token context. It is released under Apache 2.0. Mistral says it matches or beats GPT-OSS 120B on several benchmarks while producing shorter outputs.

• 119B total / 6B active MoE (128 experts, 4 active)
• 256k-token context; text and image input
• Apache 2.0 open weights
• Lab claims 3x requests per second vs Mistral Small 3

AccessMistral API and AI Studio; open weights on Hugging Face (Apache 2.0); NVIDIA build.nvidia.com and NIM; runs in vLLM, llama.cpp, SGLang, Transformers.

Price$0.15 input / $0.60 output per 1M tokens

Voxtral Transcribe 2 (Voxtral Mini Transcribe V2 and Voxtral Realtime)

Mistral's second-generation speech-to-text family, in two models. Voxtral Mini Transcribe V2 handles batch transcription with speaker labels. Voxtral Realtime is a 4B open-weights model for live transcription, with delay configurable below 200 ms. Both cover 13 languages, and Mistral states support for audio up to three hours.

• 13 languages incl. English, Chinese, Hindi, Arabic
• Speaker diarization and word-level timestamps
• Realtime: 4B parameters, delay configurable below 200 ms
• Realtime weights under Apache 2.0

AccessMistral API (both models); Voxtral Realtime also as open weights on Hugging Face under Apache 2.0. Mini Transcribe V2 is API-only.

Devstral 2

Mistral's second-generation agentic coding models, launched with the Mistral Vibe command-line coding agent. Devstral 2 (123B) scores 72.2% on SWE-bench Verified per Mistral. The 24B Devstral Small 2 is sized to run locally on consumer hardware. Mistral now presents Mistral Medium 3.5 as Devstral 2's replacement.

• 123B (Devstral 2) and 24B (Devstral Small 2) parameters
• 256K-token context window
• Lab-stated SWE-bench Verified 72.2% / 68.0%
• Small 2 deployable on consumer hardware

AccessMistral API (free at launch, then paid); Mistral Vibe CLI; open weights: Devstral 2 (123B) under a modified MIT licence, Devstral Small 2 (24B) under Apache 2.0.

PriceDevstral 2: $0.40 input / $2.00 output per 1M tokens; Devstral Small 2: $0.10 / $0.30 (after the initial free period)

Mistral Large 3

A 675B-parameter sparse mixture-of-experts model with 41B active parameters, released under Apache 2.0 as part of the Mistral 3 family. It handles text and images in more than 40 languages. At launch Mistral said it ranked second among open-source non-reasoning models on the LMArena leaderboard.

• 675B total / 41B active MoE
• Text and image input
• 40+ languages
• Apache 2.0 open weights

AccessMistral AI Studio; open weights on Hugging Face (Apache 2.0); Amazon Bedrock, Azure Foundry, IBM watsonx, OpenRouter, Fireworks, Together AI, Modal and others.

Price$0.50 input / $1.50 output per 1M tokens

Ministral 3 (3B, 8B, 14B)

A family of small dense models at 3B, 8B and 14B parameters, released alongside Mistral Large 3 under Apache 2.0. Each size comes in base, instruct and reasoning versions with image understanding, aimed at edge and low-cost use. Mistral reports 85% on AIME '25 for the 14B reasoning model.

• 3B, 8B and 14B sizes
• Base, instruct and reasoning variants
• Image understanding; multilingual
• Lab-stated 85% on AIME '25 (14B reasoning)

AccessOpen weights on Hugging Face (Apache 2.0) in base, instruct and reasoning variants; Mistral AI Studio and the same cloud partners as Mistral Large 3.

Price8B: $0.15 input / $0.15 output per 1M tokens (Mistral's example); 3B and 14B prices not captured