Mistral Medium 3.5 Is Live in Kilo Code

Key Takeaways

• 128‑billion‑parameter model with 256k context window.
• Scores 77.6% on SWE‑Bench, beating Devstral 2 and Qwen 3.5.
• Open weights released under a modified MIT license.
• Runs self‑hosted on four GPUs, available via Kilo API.
• Priced $1.50 per million input tokens, $7.50 per million output.

Pulse Analysis

Mistral’s Medium 3.5 marks a turning point for open‑source large language models, blending instruction following, reasoning and code generation into a single 128‑billion‑parameter dense architecture. In a market dominated by proprietary giants, releasing the model’s weights under a modified MIT license signals a commitment to community‑driven innovation and lowers the barrier for researchers and startups to experiment with frontier‑level AI without licensing constraints.

Technically, Medium 3.5 offers a 256k token context window, enabling long‑horizon agentic tasks such as multi‑step code refactoring, incident analysis, and complex data synthesis. Its 77.6% SWE‑Bench Verified score places it ahead of competitors like Devstral 2 and Qwen 3.5, while the vision encoder, trained from scratch, handles variable image sizes for multimodal workflows. Crucially, the model can run on as few as four GPUs, making high‑performance inference feasible for midsize teams that lack massive GPU farms.

From a business perspective, Kilo’s integration streamlines adoption: the model is accessible via the VS Code extension, CLI, cloud agents, and KiloClaw recipes, all under a single login. At $1.50 per million input tokens and $7.50 per million output tokens, the pricing is competitive for a frontier‑class model, especially when leveraged for agentic runs that capitalize on its extended context and reasoning depth. The blended pricing tier—$3 per million tokens for chat and $1.56 for long‑context summarization—further widens its appeal, positioning Medium 3.5 as a cost‑effective engine for enterprises seeking robust, open‑source AI capabilities.