Mistral AI released Medium 3.5 on April 29, its first merged flagship model — unifying instruction following, reasoning, and code capabilities in a single 128B dense model, no longer needing to maintain multiple specialized models, dramatically reducing deployment and routing complexity.
Core technical specs: 128 billion fully-activated parameters (not MoE), 256K context window, self-developed vision encoder supporting variable-resolution image input. Inference effort can be configured per request — simple Q&A takes the fast path, complex code tasks switch to deep reasoning mode, the same model covering both scenarios.
Benchmark performance is worth attention. SWE-Bench Verified reached 77.6%, very close to Gemini 3.1 Pro Preview's 78.8%. For real GitHub issue fixing tasks, Medium 3.5's performance has entered the top tier. Tau3-Telecom 91.4% demonstrates its domain-specific agent capability.
Compared to predecessors: Medium 3.1 was a pure-text model, Magistral handled reasoning, Devstral 2 specialized in code — three parallel paths for a year, now merged into one. From a product strategy perspective, this is pragmatic technical convergence: moving from model matrix to unified base, using a single configurable model to cover more scenarios.
Open-source strategy continuity is also worth attention. Medium 3.5 adopts a modified MIT license (open weights, commercial use, revenue exception clause), supports self-hosting (four H100s suffice). API pricing of $0.5/1.5 per million tokens is also lower than comparable competitors.
For developers, this means another high-cost-performance open-source option besides GPT-5.5 and Claude Opus 4.7. Especially in mid-team scenarios needing both reasoning depth and deployment cost, the "just right" nature of 128B dense models is becoming a new sweet spot.