Mistral Medium 3.5 🧠 Model Open source

What it is

Mistral AI is a Paris-based lab that publishes both hosted and open-weight models. Its current flagship is Mistral Medium 3.5 (28 April 2026), a 128B multimodal model with open weights under a modified MIT licence. The lineup also includes Mistral Large 3 (December 2025, Apache 2.0) and Mistral Small 4 (March 2026, Apache 2.0), plus specialised OCR and speech models.

Key features

• 256K-token context window on Medium 3.5, with function calling and structured outputs
• One model covers chat, reasoning and coding (Medium 3.5 replaced separate Magistral and Devstral models)
• Small 4 runs on modest hardware for private, local drafting
• Strong European-language coverage
• Batch processing at 50% off and caching that cuts input costs by up to 90%

Where you can use it

In Mistral Vibe, the assistant formerly called Le Chat (renamed 28 May 2026; our slug is still le-chat), on web, mobile and as a coding CLI. Developers use La Plateforme / Studio, and n8n has a Mistral Cloud Chat Model node. The open weights run on vLLM and other local stacks.

Pricing and rights

API: Medium 3.5 costs $1.50 input / $7.50 output per million tokens. Vibe plans: Free (limited, with $10 per month of API credits), Pro $14.99 per month with $30 of API credits (students $5.99), Team $24.99 per user. The Apache 2.0 models allow unrestricted commercial use; the Medium 3.5 licence is free for individuals and most companies, but businesses above a revenue threshold must arrange a commercial licence with Mistral.

Who it is for

European agencies and publishers with data-residency concerns, and teams who want open weights they can host.

Verdict

Mistral is the main European option with genuinely open weights and fair prices. It trails the US frontier labs on long-form writing quality, and the Medium 3.5 licence carries a revenue clause that bigger companies must check.