Mellum2.1

Mellum2.1 is an updated 12B mixture-of-experts coding model with 2.5B active parameters released under Apache 2.0; the core architecture remains unchanged from the previous version, but performance was dramatically improved through large-scale post-training reinforcement learning. Training emphasized RL as the main stage, running millions of sandboxed episodes across thousands of realistic environments and assembling tightly filtered RL datasets spanning math, competitive programming, science, tool use, and software engineering. Those changes give the model agentic capabilities: it can explore a repository, propose and edit files, and validate its own changes, making it suitable for autonomous coding tasks inside real codebases.

Benchmarks against the prior version and comparable open models (Qwen3.5-9B and Gemma 4 E4B) show the biggest gains in agentic coding and measurable improvements across coding, competitive programming, math, tool calling, and general knowledge. Because the architecture was left intact, latency remains low and multi-token prediction (MTP) improves throughput: under heavy load Mellum2.1 serves nearly twice the tokens of Qwen3.5-9B and is about 1.6× faster on single requests. Key use cases include acting as a worker inside agentic systems, tackling complex reasoning, and private self-hosted deployment. Model weights are available on Hugging Face, with GGUF builds and MTP runtime support coming soon, and user feedback is invited to guide future iterations.