Mistral Medium 3.5: The Self-Hostable Coding Model
Mistral Medium 3.5 is a 128B dense, open-weight model with a 256K context window, released at the end of April. On May 22 it became the engine behind Mistral's Vibe remote coding agents and the default model in Le Chat. It replaces two specialist models, Devstral 2 for coding and Magistral for reasoning, with one checkpoint.
The claim to test is simple: 77.6% on SWE-bench Verified from a model you can self-host on four GPUs, at $1.50/$7.50 per million tokens through the API. If that holds, Medium 3.5 is the most practical open coding model for teams that need to run models on their own infrastructure.
What it is
- 128B dense parameters: no mixture of experts. Every token uses the full model, so throughput is lower than a sparse model with the same total parameters, but memory needs are predictable.
- 256K context and configurable reasoning effort per request.
- Native vision: a vision encoder trained from scratch, handling variable image sizes and aspect ratios.
- License: modified MIT. Weights are on Hugging Face; the modification routes very large enterprises to Mistral's commercial terms, so check it if you are one.
The benchmark claims
Mistral reports 77.6% on SWE-bench Verified and 91.4 on τ³-Telecom, an agentic benchmark for multi-step tool use. It also says Medium 3.5 outperforms Devstral 2 and Qwen3.5 397B A17B. For comparison, DeepSeek reports 79.0% for V4-Flash and 80.6% for V4-Pro on SWE-bench Verified in max mode. Medium 3.5 lands just below the DeepSeek pair with a much smaller total parameter count. All of these are vendor numbers.
LMRank currently scores Medium 3.5 at 8.3.
Pricing and how it compares
| Model | LMRank | Input $/M | Output $/M | Context | Weights |
|---|---|---|---|---|---|
| Claude Sonnet 4.6 | 9.0 | $3.00 | $15.00 | 1M | Closed |
| Mistral Medium 3.5 | 8.3 | $1.50 | $7.50 | 256K | Modified MIT |
| Kimi K2.6 | 8.4 | $0.95 | $4.00 | 262K | Modified MIT |
| Devstral 2 | 8.3 | $0.40 | $2.00 | 262K | Open |
| Qwen3.6 35B A3B | 8.5 | $0.15 | $1.00 | 262K | Apache 2.0 |
Through the API, Medium 3.5 is neither cheap nor expensive: half the price of Sonnet 4.6, but several times the price of Kimi K2.6 or Qwen3.6 35B A3B, which score about the same on LMRank. It also costs more than the Devstral 2 it replaces. On price alone the API case is weak. The reason to choose it is self-hosting.
Why the self-hosting story matters
A 128B dense model fits on four modern data-center GPUs. That is a normal purchase for a mid-size engineering organization, unlike the multi-node clusters needed for trillion-parameter MoE models like DeepSeek V4 Pro or Kimi K2.6. For teams that can't send code to an external API, whether for regulatory or contractual reasons, Medium 3.5 is one of the strongest coding models they can realistically run in-house, and it comes from a European vendor, which matters for some procurement rules.
The trade-off is throughput. A dense 128B model generates tokens more slowly per GPU than a sparse model with 10B to 30B active parameters. For interactive coding, that shows up as latency; for background agents, it shows up as GPU-hours.
Takeaway
If you can use a hosted API, Kimi K2.6 or Qwen3.6 35B A3B give similar LMRank scores for much less. If you need to self-host a capable coding agent on a four-GPU budget, Mistral Medium 3.5 is the model to evaluate first. Run it on your own repositories; a 77.6% SWE-bench claim tells you little about your codebase.
Related: Best Agentic Coding Models · Best Open-Weight Models