Mistral released Medium 3.5 on April 29, 2026, and the most interesting thing about it isn't a benchmark number — it's the strategy. Where every other open-weight flagship in 2026 has gone Mixture-of-Experts, Mistral Medium 3.5 is dense, 128 billion parameters, with a 256K context window. And it consolidates three previously separate product lines into a single checkpoint.
Mistral Medium — general-purpose instruction following and reasoning
Devstral — Mistral's coding-focused line
Pixtral — Mistral's vision-language line
In the new architecture, all 128 billion parameters activate on every token — not the MoE pattern of "17B active, 400B total." This is a deliberate counter-bet against the MoE trend, betting that dense models still produce more consistent quality per active parameter, especially for reasoning across long contexts.
The release also ships:
256K context window
Configurable reasoning effort in a single set of weights — closer to the Anthropic / OpenAI thinking-vs-instant pattern
Multimodal input (text + images) baked into the same checkpoint
A modified MIT open-weights license — permissive enough for most commercial use, with a few enterprise carveouts in the Mistral terms
The Vibe agent
The release came alongside Mistral Vibe, Mistral's agentic coding product. Vibe runs Medium 3.5 as the underlying model and ships:
Remote cloud agents — long-running execution outside your laptop
A Vibe CLI that opens PRs against your repos directly
Integration with Le Chat's new Work mode — same checkpoint, different surface
Per the release notes, Mistral Medium 3.5 supersedes Devstral 2 in Vibe — the previous coding model was effectively replaced by the new merged checkpoint, which scored higher on agentic coding benchmarks across the board.
Benchmarks at launch
The numbers Mistral published:
τ³-Telecom: 91.4% — a multi-turn agentic-task benchmark
SWE-Bench Verified: 77.6% — solid mid-tier on the most-watched coding benchmark, behind Opus 4.7 (87.6%) and GPT-5.5, but ahead of most open-weight competitors
[Inference] On general-knowledge and reasoning benchmarks, Medium 3.5 sits roughly where you'd expect a 128B dense model — competitive with frontier on knowledge-bound tasks, slightly behind on the hardest reasoning evals. The unified positioning is more valuable than peak benchmark scores: one model that does coding, vision, and reasoning is operationally simpler than three.
Why dense, in 2026
The case for sticking with dense in 2026 is non-obvious but specific:
Per-active-parameter quality — A dense model with 128B active beats an MoE with 17B active on equal-active-FLOPs comparisons in many academic benchmarks. The MoE bet is that you can compensate with bigger total capacity; the dense bet is that consistent specialization beats router-driven specialization on hard tasks.
Operational simplicity — Dense models are easier to fine-tune, serve, and reason about. No expert-parallel sharding, no router-collapse failure modes, no load-balancing tuning.
More predictable latency — MoE routing adds variance. Dense doesn't.
Memory cost is honest — A 128B dense model needs ~256GB at full precision (or ~64GB at 4-bit quantization). That's a known, fixed cost. MoE models have hidden costs in expert parallelism that don't show up in the parameter count.
The case against: at 128B dense, you're paying full FLOPs per token, every token. Llama 4 Maverick at "17B active" runs at materially lower cost per token despite being a larger total model.
What 256K context buys you
256K tokens is shorter than Claude Opus 4.7 (1M) or Gemini 3.1 Pro (1M), longer than most older Mistral models. For most real-world workloads, 256K is enough:
Full medium-sized codebase in context
A book-length document with annotations
Multi-file diff review with surrounding tests
The cases where 256K is too short — repository-scale agentic coding, long-running research synthesis — are also the cases where retrieval-augmented patterns work better than naive full-context prompts. So the practical-vs-marketing gap on context windows is smaller than it looks.
Pricing and availability
Mistral has historically priced Medium aggressively against the proprietary leaders, and 3.5 follows the pattern. The model is available:
Via the Mistral API (proprietary endpoint)
Via the modified MIT-licensed open weights on Hugging Face
Inside Le Chat's Work mode and the Vibe agent
For self-hosters, the dense-128B form factor is operationally friendlier than running a 400B MoE — fewer GPUs, simpler tooling, easier fine-tuning. For API users, the question is whether you'd rather pay a frontier-tier rate for Opus 4.7 or GPT-5.5, or a more moderate rate for a model that lands a notch below them on most evals.
Who Medium 3.5 is for
The clearest fits:
European or sovereignty-constrained deployments — Mistral's positioning as an EU-headquartered alternative still matters for regulated buyers
Self-hosters who want frontier-adjacent quality without MoE operational overhead
Teams consolidating multi-model stacks — if you're running Devstral + Pixtral + Medium today, this is a one-checkpoint replacement
Coding agents that don't need the absolute top of the SWE-Bench leaderboard — Vibe's pitch is that 77.6% verified SWE-Bench plus a great agent loop beats 87% verified plus a worse agent
The bottom line
Medium 3.5 is the most credible "dense flagship" in 2026 — a deliberate counter-bet against the MoE trend, a unified replacement for three previously separate Mistral product lines, and the foundation under a new agentic coding stack. It probably won't top your benchmarks. It might top your operational simplicity, your sovereignty story, and your fine-tuning workflow.
Frequently Asked Questions
When was Mistral Medium 3.5 released and what's its license?
Medium 3.5 was released on April 29, 2026, as a 128 billion parameter dense model with a 256K context window. It's distributed under a modified MIT open-weights license, permissive enough for most commercial use with some enterprise-specific terms.
How does Mistral Medium 3.5 compare to Llama 4 Maverick?
Maverick is a 400B-total / 17B-active MoE model with a 1M context; Medium 3.5 is a 128B dense model with a 256K context. Maverick has more total knowledge capacity and lower per-token compute; Medium 3.5 is operationally simpler, more predictable in latency, and easier to fine-tune. They target different deployment shapes.
What is Mistral Vibe and how does it use Medium 3.5?
Mistral Vibe is Mistral's agentic coding product, including a CLI that opens PRs against your repositories and remote cloud agents that execute long-running tasks. Vibe runs Medium 3.5 as the underlying model, replacing the previous Devstral 2 model on every benchmark per the launch notes.
