Mistral Medium 3.5 arrives in Kilo Code preview
Kilo Code has added public preview support for Mistral Medium 3.5, Mistral’s new 128B blended model. The company says it is available across the Gateway, VS Code extension, CLI, cloud agents, and KiloClaw recipes.
Kilo Code has added public preview support for Mistral Medium 3.5, Mistral’s new blended model built for instruction-following, reasoning, and coding. Kilo said the model is now available across its product line, including the Gateway, VS Code extension, CLI, cloud agents, and KiloClaw recipes.
According to Kilo, Mistral Medium 3.5 is Mistral’s first blended model, combining instruction-following, reasoning, and coding into a single 128B dense model. The company said it was built for long-horizon agentic work and includes a 256k context window, which gives it room to work with large prompts and long sessions.
⚡ New to this?
This news is about a new AI model from Mistral and where it can be used inside Kilo Code. A blended model means one system is built to handle multiple jobs at once, such as following instructions, reasoning through problems, and writing code.
For non-experts, the main point is that this model is meant for longer, more complicated AI tasks, especially in software work. The “context window” is how much text or code the model can keep in mind at once, so a larger one helps with big files, long chats, and multi-step jobs.
🦞 OpenClaw angle
If you run agent workflows, try this model first for long refactors, CI troubleshooting, and multi-file code changes where context loss hurts results. Use it in the cloud agent setup when you want the model to work asynchronously and return a branch or PR draft instead of a short answer.
If you self-host, note the four-GPU claim and test whether your hardware can support the 128B model before putting it into a production queue. For automation pipelines, set it up as the default only for tasks that need long context or heavier reasoning, and keep a cheaper model for short edits and routine chat.
Kilo also said the model includes configurable reasoning effort, allowing users to turn up the depth for harder tasks such as major refactors or keep it lighter for smaller edits. The company said the model can run self-hosted on as few as four GPUs.
On benchmark results, Kilo cited a 77.6% score on SWE-Bench Verified. The company said that places it ahead of Devstral 2 and models including Qwen3.5 397B A17B.
The model also includes a vision encoder trained from scratch to handle variable image sizes, according to Kilo. The company said Mistral is shipping the model with open weights under a modified MIT license.
Mistral Medium 3.5 is now the default model for Mistral Vibe CLI and Le Chat, according to the company. Kilo said the new model is also available through the Kilo Gateway, which lets users access it with a single login alongside other models.
In the VS Code extension, Kilo said the updated model switcher now includes Mistral Medium 3.5. The company described it as a fit for tasks that need large context, step-by-step reasoning, and structured output that codebases can use directly.
Kilo said the CLI version is aimed at longer terminal sessions such as dependency upgrades, test generation, and CI investigations. The company said the model is also available in cloud agents for async work, including module refactors, issue triage, test coverage gaps, and incident investigations.
Kilo added that Mistral Medium 3.5 can be used across KiloClaw recipes, for both personal and work workflows. The company said it is priced at $1.50 per million input tokens and $7.50 per million output tokens through the API.
For general chat, Kilo said the blended price is $3 per million tokens, while long-context summarization is priced at $1.56 per million tokens. The company said the model is available now in public preview.
