Mistral Medium 3.5 lands as a 128B dense model with agentic features

Mistral AI released Medium 3.5 on April 29, a dense 128B-parameter model with a set of new agentic features [Mistral models].

── What shipped ──

Medium 3.5 fits into the Mistral 3 family released in late 2025/early 2026, which already includes Mistral Large 3 (sparse MoE, 41B active / 675B total, Apache 2.0) plus dense Small 4, 8B, and 3B variants.

The agentic features were not detailed in the release notes beyond high-level capability claims around tool use, planning, and multi-step task execution. Pricing and benchmarks were published alongside the model card [Mistral docs].

── Why it matters ──

Mistral has held to a steady open-weight release cadence — Voxtral TTS in March, Leanstral for Lean 4 proof engineering in March, Small 4 in March, Medium 3.5 in April. While the American frontier has consolidated around closed-weight, billion-dollar-compute models, Mistral keeps shipping.

The competitive position is clear: not first on capability, but first on practical deployability. A 128B dense model is large enough to be capable on real tasks and small enough to self-host on enterprise hardware without exotic infrastructure.

For European companies subject to data-residency rules, Mistral remains the only credible option that combines frontier-adjacent capability with provider-neutral on-prem deployment.

── Editor's take ──

The interesting Mistral story is not Medium 3.5 itself. It is whether Mistral can sustain a release cadence on the open-weight track while their American peers concentrate on closed-weight scaling. The bet: enterprise on-prem buyers will pay for "good enough + you own it" over "best + we own it." Worth watching.