Mistral Medium 3.5: 128B dense open-weights, 77.6% SWE-Bench Verified

Mistral pushed Medium 3.5 to Hugging Face this week with weights open under their license: 128B dense parameters, 256k context, 77.6% on SWE-Bench Verified, 91.4 on τ³-Telecom. The thing that matters for builders running self-hosted agents is the combination — coding-capable backbone you can pull, fine-tune on your codebase, and serve on your own GPUs. The closed frontier still leads, but the gap on long-tail real-issue resolution has compressed enough that hosting choices come back into play.

Two architectural choices to flag. First, dense not mixture-of-experts: Medium 3.5 beats Qwen3.5 397B-A17B (MoE, ~17B active) at SWE-Bench despite being smaller in absolute weights. The "merged model" language Mistral uses means they collapsed the prior split between Mistral instruct and Devstral coding-specialist into a single weight set covering instruct, reasoning, and coding — simpler ops for builders who hated juggling two endpoints. Second, the 77.6% number is single-pass on the 500-task Verified subset; Sonnet 4.5's 82% came with parallel test-time compute, so the right comparison is closer than the headline suggests. What Mistral didn't disclose is the contamination story or whether the Vibe harness post-processes — that's the next question to ask before anyone ports Medium 3.5 into a production loop.

The Vibe surface is the other half of this release. Vibe was already Mistral's CLI coding agent — same category as Claude Code, Cursor's Composer, Aider — but Remote Agents make it a proper Cursor/Devin competitor: sandboxed cloud execution of long-running tasks while you work elsewhere, sessions launchable from CLI or Le Chat. The ecosystem read is that open-weights labs are no longer just shipping the model and leaving the agent surface to wrappers. Mistral is closing the loop themselves, the way Anthropic shipped Claude Code alongside Sonnet 4.5. For builders, this means the open stack is now end-to-end credible: weights you can host, agent surface you can use directly, or peel off and integrate piecewise. The closed labs' moat narrows to test-time compute, deeper tool integration, and whatever the CAISI pre-release eval pipeline confers.

Practical move: if you're running Devstral 2 or a non-Mistral coding specialist behind your agent, Medium 3.5 is worth a benchmark swap on your own eval set this week. Single weight set simplifies the deploy, 256k context handles real-codebase windows, and Vibe Remote Agents are usable out-of-the-box if you don't want to build sandboxing yourself. If you're already on closed-frontier API and watching per-token economics, a 128B dense model is small enough to make self-hosting math work on a single 8xH100 node — that's the calculation that's been missing for the open-weights agent pitch.