Mellum2.1 moves its training weight to reinforcement learning: the 12B open model now explores repos, edits files and checks its own changes
Synopsis
Mellum2.1 is the next version of Mellum2, with the architecture unchanged: still a 12B mixture-of-experts model with 2.5B active parameters under Apache 2.0, with the changes concentrated after pre-training. Mellum2 was fast but could not work inside a repository at the level its developers wanted; this version, trained with reinforcement learning in real environments across thousands of environments and millions of sandboxed runs, can explore a codebase, edit files and check its own changes. Under one shared evaluation setup it is compared with Mellum2 and with the similar open models Qwen3.5-9B and Gemma 4 E4B, and the largest gain is in agentic coding, with improvements also reported in coding, competitive programming, math, tool calling and general knowledge.
Interpretation
Mellum2.1 shifts training weight from a short post-pre-training stage to reinforcement learning, so the 12B mixture-of-experts model can explore a repository, edit files and verify its own changes. The previous Mellum2 was fast but could not work inside a repository at the level its developers wanted; every change in this version happens after pre-training, while the architecture and the 2.5B active-parameter scale stay the same. The evaluation compares Mellum2.1 with Mellum2 and with the similar open models Qwen3.5-9B and Gemma 4 E4B under the same evaluation setup, and agentic coding is the area with the largest improvement.
On the training side, reinforcement learning runs in real environments, and the model repeatedly meets executable, verifiable tasks during training. Reinforcement learning goes from a short final stage to the main part of training, with new tasks added in math, competitive programming, science, tool use and software engineering, mixing open reinforcement-learning datasets with tasks built in-house. Open data often carries broken tests, unverifiable answers, or tasks that are too easy or impossible, so every source is filtered before it reaches training; the supporting infrastructure runs thousands of reinforcement-learning environments in-house and launched millions of sandboxes over the course of training.
Speed does not drop because of post-training, and multi-token prediction makes a single request faster. Post-training leaves the architecture untouched, so Mellum2.1 is as fast as Mellum2, and multi-token prediction adds further speed on top. Under heavy load Mellum2.1 is the fastest model in the group and serves almost twice as many tokens as Qwen3.5-9B; for a single request, multi-token prediction makes it about 1.6 times faster.
Perspective
The demonstrated scope is coding agents and sub-agents: exploring a repository, editing files, checking changes, plus tasks in math, competitive programming, tool calling and general knowledge; the model is available on Hugging Face for users running it locally or on their own infrastructure. GGUF builds for llama.cpp, Ollama and LM Studio, and the multi-token prediction head for speculative decoding in vLLM, are described as coming soon.
The public material describes performance gains qualitatively and gives no per-benchmark scores; being fastest under heavy load, serving almost twice the tokens, and the roughly 1.6 times single-request speedup come without specific measurement conditions. The GGUF builds and the multi-token prediction head for vLLM speculative decoding are not yet released.