Mistral AI
European lab publishing open weight models you can run on your own hardware. Large 3 is a 675B MoE with 41B active, 256K context and Apache 2.0 weights.
Overview
Mistral is the European answer to what happens when prompts cannot leave the continent. Large 3 arrived in December 2025 as a sparse mixture of experts, 675B total parameters with 41B active on any given token, a 256K context window and published weights. Small 4 and Medium 3.5 sit below it, Codestral handles code and the Ministral family covers edge sized deployment. Most of the lineup ships under Apache 2.0, with Medium 3.5 on a modified MIT licence.
The hosted API is aggressive on price. Large 3 runs $0.50 per million input and $1.50 output, with Small 4 at $0.15 and $0.60. The user-facing surface, Le Chat until 28 May 2026 and Vibe since, moved from chat assistant to agent platform: Work Mode for long-horizon tasks, Code Mode for remote coding in isolated sandboxes, a VS Code extension, a terminal CLI and cloud Remote Agents that hand back a pull request. Pro is $14.99 a month, Team $24.99 per user, students $5.99.
Worth evaluating if you carry GDPR or sovereignty requirements, if you want to fine tune weights you control through Forge or if you are building a second supplier so a single US vendor cannot price or policy you into a corner. Skip it if the goal is topping a coding leaderboard, because Mistral is honestly a tier down there. Self-hosting also needs a real cost model rather than an instinct: 41B active parameters on a 675B model demands meaningful GPU capacity, and at $0.50 per million on the hosted API you have to run a lot of tokens before owning the hardware pays back.
Key Features
- ✓ Mistral Large 3, a 675B parameter sparse mixture of experts with 41B active per token and a 256K context window, released December 2025
- ✓ Open weights across most of the lineup, Apache 2.0 for Large 3 and Small 4, modified MIT for Medium 3.5
- ✓ Self-hosted, hybrid and in-VPC deployment with no per-token cost and no prompts leaving your network
- ✓ Vibe, the product renamed from Le Chat on 28 May 2026, running Work Mode and Code Mode across a VS Code extension, a terminal CLI and cloud Remote Agents that deliver through to a pull request
- ✓ Mistral Studio for building and evaluating agents, Forge for fine tuning and alignment, plus Mistral OCR 4 billed per thousand pages and Voxtral TTS billed per minute
Where it holds
- • Open weights at near-frontier quality is a genuinely different offer. If your compliance position is that model weights must sit on your own hardware, the shortlist is very short and this is on it.
- • Large 3 at $0.50 input and $1.50 output on the hosted API undercuts every rival flagship by a wide margin, roughly a twentieth of Sol's output rate.
- • EU data residency and AI Act aligned documentation ship as standard rather than as an enterprise upsell.
Where it breaks
- • Raw capability trails. On general reasoning and hard agentic coding it sits behind GPT-5.6 Sol and Claude Opus, and Mistral does not seriously contest that.
- • The May 2026 rename of Le Chat to Vibe threw away real brand recognition, and docs, third party guides and search results still disagree on what to call things.
- • Self-hosting a 675B MoE is not free in practice. 41B active parameters still needs serious GPU capacity, and at $0.50 per million hosted the break-even sits much further out than teams expect.
- • Licensing is not uniformly Apache 2.0. Check the specific model before assuming commercial self-hosting carries no obligations.
My Take
It will not win a benchmark shootout. On general reasoning and agentic coding, Large 3 sits behind GPT-5.6 Sol and Claude Opus, and pretending otherwise wastes time. Where it wins is the deployment question: a 675B mixture of experts with 41B active parameters, a 256K window and Apache 2.0 weights you can pull down and run inside your own network with no per-token bill. For a manufacturing or defence-adjacent client holding a hard data residency line, that often settles the evaluation before capability is even discussed. The rename of Le Chat to Vibe in May 2026 was a brand mistake and the documentation still has not fully caught up.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0
- Starting at:
- Open weights are free to self-host: Large 3 and Small 4 under Apache 2.0, Medium 3.5 under a modified MIT licence. Hosted API per million tokens: Mistral Large 3 $0.50 in / $1.50 out, Small 4 $0.15 / $0.60, Codestral $0.30 / $0.90, Ministral 3 14B $0.20 / $0.20. Batch takes 50% off and cached input up to 90% off. Vibe plans: Free with $10/mo API credits, Pro $14.99/mo with $30/mo credits, Student $5.99/mo, Team $24.99/user/mo, Enterprise custom.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Mistral AI holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Mistral AI
DeepSeek
freemiumChinese AI lab producing open-source LLMs that rival proprietary models at a fraction of the cost, with strong reasoning and coding abilities.
Google Gemma 4
open-sourceGoogle's open-weight family in five sizes, 2B effective up to 31B dense, text and image in with audio on the smaller ones, 256K context and Apache 2.0 weights since April 2026.
NVIDIA Nemotron 3
open-sourceNVIDIA's open-weight family, 31.6B Nano up to the 550B Ultra MoE. Hybrid Mamba-Transformer, 1M context, weights plus roughly 3T tokens of training data published.
MiniMax M3
freemiumMiniMax's open-weight flagship LLM (稀宇科技), distinct from the Hailuo video product: 428B mixture-of-experts, 23B active per token, 1M context and native image and video input.