Mistral Medium 3.5: 128B Open-Weight Agentic Coder
On May 22, Mistral AI shipped a model that barely registered on the hype cycle - and that might be the most interesting thing about it. Mistral Medium 3.5 is a dense 128B open-weight model with a 256K context window, built specifically for long-horizon agentic coding and productivity work. It replaces Devstral 2 in Mistral's Vibe CLI, becomes the default model in Le Chat, and ships under a modified MIT license that permits commercial use and self-hosting on as few as four GPUs.
It also claims some of the strongest coding benchmark numbers we've seen from a sub-200B parameter model: 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom. Those aren't merely good numbers. They're "why am I paying flagship prices for coding agents?" numbers.
What Mistral Medium 3.5 Actually Is
Mistral calls Medium 3.5 their "first flagship merged model" - a single dense checkpoint that handles instruction-following, reasoning, and coding without routing between specialist heads. At 128B parameters, it sits in a size class that can run on consumer-grade server hardware while delivering performance that punches above its weight.
The headline specs:
- 128B dense parameters - Not MoE, so every forward pass uses the full weights. That means higher latency than a 3B active-parameter model, but also more consistent quality on complex reasoning.
- 256K context window - Matching Qwen3.6 35B A3B and exceeding Claude Sonnet 4's 200K context.
- Open weights - Modified MIT license, commercially usable, self-hostable.
- Configurable reasoning effort - The same checkpoint can answer a quick chat reply or work through a complex multi-step agentic run.
- Native vision - A vision encoder trained from scratch handles variable image sizes and aspect ratios.
The model is currently in public preview and available via Mistral's API, Le Chat, and OpenRouter.
The Benchmark Claims
Mistral reports Medium 3.5 scores 77.6% on SWE-Bench Verified, a rigorous benchmark for real-world software engineering. For context, that puts it ahead of the previous Qwen3.6 35B A3B claims (which underpins an 8.5 LMRank score) and well ahead of Codestral (7.8 score). It also scores 91.4 on τ³-Telecom, an agentic capability benchmark measuring reliable multi-tool calling across complex tasks.
On the LMRank leaderboard, Mistral Large 3 currently holds an 8.3 score. Medium 3.5 has not yet completed independent evaluation on our methodology, but if Mistral's SWE-Bench and τ³ claims hold up in third-party testing, it could compete in the same tier as GPT-5.5 Instant (8.6) and Claude Sonnet 4 (8.8).
Pricing and the New Medium Tier
On OpenRouter, Mistral Medium 3.5 is priced at $1.50 per million input tokens and $7.50 per million output tokens. Here's how that compares to coding-focused models already on the LMRank leaderboard:
| Model | Score | Input $/M | Output $/M | Context |
|---|---|---|---|---|
| Claude Sonnet 4 | 8.8 | $3.00 | $15.00 | 200K |
| GPT-5.5 Instant | 8.6 | $2.00 | $10.00 | 128K |
| Mistral Medium 3.5 | TBD | $1.50 | $7.50 | 256K |
| Kimi K2.6 | 8.4 | $0.95 | $4.00 | 256K |
| Qwen3.6 35B A3B | 8.5 | $0.15 | $1.00 | 262K |
The pricing is competitive without being the cheapest. At $1.50 per million input tokens, Medium 3.5 undercuts Claude Sonnet 4 by 50% and GPT-5.5 Instant by 25%. It costs more than Kimi K2.6 and Qwen3.6 35B A3B, but both of those are MoE models with smaller active parameter counts. If you need dense-model consistency - especially for agentic workflows where reliability matters more than raw speed - Medium 3.5's price is justified.
Why Open Weights Change the Game
The most important detail about Medium 3.5 isn't the benchmark score or the context window. It's the license. Modified MIT means you can download the weights, run them on your own hardware, and build products on top of them without negotiating an enterprise deal.
For a 128B dense model, self-hosting on four GPUs is genuinely practical for mid-size engineering teams. That's a very different proposition from Llama 4 405B, which requires a small data center, or Qwen3.7 Max, which is API-only. Medium 3.5 occupies a rare sweet spot: frontier-adjacent capability with local-deployment feasibility.
The Bottom Line
Mistral Medium 3.5 is the clearest sign yet that the "medium" tier is becoming the most interesting battlefield in AI. Flagship models like Claude Opus 4.8 and GPT-5.5 still win on absolute reasoning depth, but they're overkill - and overpriced - for the bulk of real-world coding work. Medium 3.5 offers a credible alternative: strong benchmark claims, open weights, a huge context window, and pricing that doesn't require CFO approval.
We're watching independent benchmarks closely. If Medium 3.5 validates its claims on LiveBench, SWE-Bench, and LMSYS arena evaluations, it will earn a spot on the LMRank leaderboard as one of the best value propositions in coding AI.
Models to watch: Mistral Large 3 (8.3) · Claude Sonnet 4 (8.8) · Kimi K2.6 (8.4) · Qwen3.6 35B A3B (8.5) · GPT-5.5 Instant (8.6). Compare pricing, context windows, and scores on the LMRank leaderboard.
At lmrank.com, we track live scores and pricing for every major LLM. Follow the leaderboard for updated rankings as Mistral Medium 3.5 and other new models complete independent evaluation.
See also: Best Agentic Coding Models