Mistral Medium 3.5

Mistral Medium 3.5 is Mistral AI’s flagship, released at the end of April 2026 (sources give April 29 to May 1). Mistral calls it its first “merged” flagship: it folds the company’s separate reasoning model (Magistral) and coding model (Devstral 2) into a single set of weights that also handles images.

Key specs

• Architecture: 128-billion-parameter dense model (not mixture-of-experts)
• Context window: 256,000 tokens
• Licence: modified MIT, open weights
• API price: $1.50 per million input tokens, $7.50 per million output
• Reasoning: configurable effort per request
• Hardware: Mistral says it can be self-hosted on as few as four GPUs

Why it matters

Most frontier-class open models in 2026 are huge mixture-of-experts systems that are hard to run. Medium 3.5 is designed to be practical to host on a company’s own servers. That appeals to European governments and regulated industries pursuing sovereign SI. Mistral reports a 77.6% score on the SWE-Bench Verified coding benchmark.