MiniMax-M2.1 is available free in Kilo for a limited time
Kilo has added MiniMax-M2.1 to its roster of hosted coding models and made it available for free in Kilo Code starting Jan 05, 2026. The model, developed by MiniMax, is an open-weight Mixture-of-Experts release that has attracted strong attention across developer communities and model hubs.
What the model brings to coding workflows
MiniMax-M2.1 ships with a 230 billion-parameter MoE architecture with only 10 billion parameters active at inference, combining large nominal capacity with runtime efficiency. It also offers a massive token context window, which supports longer, multi-step development tasks and agentic workflows such as complex debugging and chained reasoning. MiniMax has published the model on Hugging Face, where it recorded almost 200,000 downloads: https://huggingface.co/MiniMaxAI/MiniMax-M2.1
Key technical highlights
- 230B parameters (MoE) with 10B active — aims to balance expressivity and efficient inference.
- Large context window — suited to long-form code, multi-file reasoning, and agent workflows.
- Multilingual capability — positioned for global developer communities.
Performance and reception
Early adopters and benchmark runs place M2.1 favorably against top proprietary coding models. Reports indicate coding performance that rivals models such as Claude 3.5 Sonnet and Gemini 3 Pro, especially for web and mobile development tasks. Public benchmark results on VIBE and SWE-bench, along with comparative posts from Code Arena, have helped fuel interest: https://x.com/arena/status/2006772410004250845
Kilo’s internal testing, covered in a deeper analysis, highlights the model’s strength in agentic scenarios and multi-step problem solving: https://blog.kilo.ai/p/minimax-m21
How it fits into Kilo and the wider OSS landscape
MiniMax is phasing out the free tier of its predecessor, M2, and M2.1’s limited-time free access in Kilo lowers the barrier for experimentation with an open-weight frontier model. The release aligns with broader trends toward higher-quality open models and increased parity with proprietary alternatives — a shift Kilo has discussed in its LLM convergence analysis: https://blog.kilo.ai/p/the-llm-convergence-threshold-has
Kilo recommends the model for tasks such as refactoring legacy code, fast code review runs, and building new projects in its App Builder: https://blog.kilo.ai/p/kilo-code-weekly-product-roundup-a13. Users can access MiniMax-M2.1 through Kilo’s platform: https://kilo.ai/ and view model rankings on Kilo’s leaderboard: https://kilo.ai/leaderboard
The MiniMax-M2.1 addition reinforces Kilo’s stated commitment to hosting a range of free, high-performance coding models for developers to test and integrate into real development workflows.
Original source: https://blog.kilo.ai/p/minimax-m21-is-free-in-kilo-for-a
