Mistral

The generalist models, with the licence each one carries:

• Mistral Large 3 (v25.12) — the flagship. A sparse mixture-of-experts with 41B active parameters out of 675B total, trained from scratch on 3,000 NVIDIA H200 GPUs, with image understanding and multilingual coverage. Apache 2.0, base and instruction-tuned versions both.
• Mistral Medium 3.5 (v26.04) — a dense 128B model with a 256K context window that merges instruction-following, reasoning and coding into one set of weights, with per-request reasoning effort. Modified MIT licence.
• Mistral Small 4 (v26.03) — a hybrid instruct/reasoning/coding model for cost-sensitive work. Apache 2.0.
• Ministral 3 at 14B, 8B and 3B (v25.12) — dense edge models, each shipped in base, instruct and reasoning variants, all with image understanding. Apache 2.0.

Around them sit the specialists: Codestral for code completion, Codestral Embed and Mistral Embed for retrieval, the OCR 4 line for document parsing, Voxtral for transcription and speech, Shieldstral 1.0 for moderation, and Leanstral for Lean 4 theorem proving. These are a mix — Shieldstral and Leanstral are Apache 2.0, Voxtral TTS is CC BY-NC 4.0 and therefore off-limits commercially, and Codestral, OCR and the embedding and moderation models are Premier-licensed and API-only.

One recent change is worth flagging: since August 2026 Mistral also serves a third-party open model, Z.ai's GLM-5.2, on its own infrastructure, under the same regional controls as its own models. The pitch is that you can widen model choice without fragmenting where inference runs.

Mistral's marketing leans on "open", and for the current generalist line that word is doing real work. Apache 2.0 on Large 3 and Ministral 3 means commercial deployment, modification, redistribution and derivative works, with no revenue threshold and no separate agreement. Medium 3.5's modified MIT licence is close to the same posture. That is a different proposition from vendors who publish a previous generation while keeping the current flagship closed.

Practical consequences follow. Large 3 ships in an NVFP4 checkpoint built with llm-compressor, which Mistral says runs on a single 8×A100 or 8×H100 node under vLLM, or on Blackwell NVL72 systems. Medium 3.5 is small enough to self-host on as few as four GPUs, per Mistral's own release notes. Ministral 3 targets DGX Spark, RTX laptops and Jetson devices. So "open weights" here is not a gesture at a checkpoint nobody can serve.

Note one wrinkle: Mistral's own pricing FAQ still carries older language saying open-weight models are "Apache 2.0 licensed for research/individual use" with commercial deployments requiring a Mistral licence. The Mistral 3 announcement and the per-model licence tags in the docs both say Apache 2.0 outright. If you are making a legal decision, read the licence file attached to the specific model on Hugging Face rather than the marketing copy.

The consumer surface is mid-rename and it is confusing. Le Chat is the chat interface at chat.mistral.ai, and Le Chat is still the name on the iOS and Android apps. Vibe is the agent brand: a terminal CLI, VS Code and JetBrains plugins, and since May 2026 remote coding agents that run in isolated cloud sandboxes, open pull requests on GitHub, and can be handed off — Mistral calls it "teleport" — from a local session. The pricing page now sells the consumer plans under the Vibe name. La Plateforme, the developer console, has been renamed Mistral AI Studio.

The May 2026 release also added a "Work mode" in Le Chat, an agentic mode with connectors on by default that runs multi-step research and cross-tool tasks and asks for explicit approval before sending messages or modifying data. Mistral Medium 3.5 is the default model behind both Work mode and the Vibe coding agent, and Mistral reports it at 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom — vendor-published figures, not independently replicated.

Free gets you the chat product on web and mobile with limits Mistral describes only as "limited" for messages, web searches and coding sessions, plus image generation, 100+ connectors, up to five scheduled tasks and $10 a month of Studio API credits. Pro is $14.99 a month excluding tax — $5.99 for verified students for up to twelve months — and raises the ceilings across the board, adds all-day coding in the CLI, IDE and web, 15GB of library storage and $30 a month in API credits. Team is $24.99 per user per month plus a $50 monthly base fee, and adds 30GB per user, domain verification and data export. Enterprise is quoted, and covers private and on-premises deployment, custom models, SAML SSO, audit logs and white labelling.

API billing is separate from all of that. As of August 2026 the published rates are $0.15 per million tokens in and out for Ministral 3 8B, $0.50 in and $1.50 out for Mistral Large 3, and $1.50 in and $7.50 out for Mistral Medium 3.5. Batch processing halves the price; cached input tokens cut input cost by up to 90%. OCR is billed per 1,000 pages, speech per minute, and tool APIs per call.

Mistral Regional Endpoints reached general availability in August 2026 and let customers pin inference to Europe or the US to match data-residency and latency requirements, with limited safeguarded sub-processor transfers documented in the Trust Center. A Priority Tier in public preview adds committed rate limits and an uptime SLA. Mistral claims to be the only European lab offering both processing-region choice and an SLA-backed service level — a claim about its competitors that is worth verifying against your own shortlist.

Self-host if you have GPUs, a compliance requirement that rules out third-party inference, or a fine-tuning programme whose output you need to own. Ministral 3 at 3B or 8B is the obvious starting point for edge and on-device work; Medium 3.5 is the interesting middle if you can find four GPUs. Use the API — or one of the partner routes on Amazon Bedrock, Azure Foundry, IBM watsonx, OpenRouter, Fireworks or Together — if you want the same models without the operational burden. Take the Pro subscription if you mainly want the chat and coding agent and the API credits are a bonus rather than the point.

The honest summary is that Mistral does not win the leaderboard argument. Its own Mistral 3 post placed Large 3 at #2 among open non-reasoning models and #6 among open-source models overall on LMArena, which is a respectable placement rather than a frontier claim. What it wins is the ownership argument, and for a large enterprise deciding where a decade of fine-tuning investment will live, that is often the argument that decides the contract.