JetBrains Releases Open-Weight Mellum2.1, Boosting Coding Agents with Reinforcement Learning in Real Repositories

JetBrains released Mellum2.1 Thinking on October 8, targeting coding agents that can explore repositories, edit files, and run tests. It retains Mellum2’s mixture-of-experts architecture, with 12 billion total parameters and about 2.5 billion active per token. The weights are available under the Apache 2.0 license. This update focuses on post-training, teaching the existing model a more complete software engineering workflow. Official announcement

JetBrains says it ran millions of sandboxed executions across thousands of environments during training, and expanded its programming, mathematics, and tool-use tasks. For software engineering training, the model worked in real repositories using shell and file-editing tools, with passing tests as the reward. The team also filtered out data with broken tests, unverifiable answers, or unsuitable difficulty, making data and execution-environment quality central to the update. Training methodology

The official model card reports that the SWE-bench Verified solve rate rose from 2.0% for the predecessor to 47.0%, while Qwen3.5-9B scored 50.0% in the same evaluation. JetBrains measured all these results itself. Agent tests consistently used Pi v0.73.1, shell and file tools, a 114K context window, and up to 16K output tokens per turn. The predecessor was also reevaluated, so the scores should be compared within this setup; comparisons across leaderboards require checking the agent tools and inference budget. Evaluation and settings

For deployment, official GGUF quantized files are available. The Q4_K_M version is about 8.1 GB and can be used with runtimes such as llama.cpp and Ollama. That figure is only the weight file size; actual memory use must also account for the KV cache and runtime overhead. Quantization quality figures come from short-context Wikitext-2 logits comparisons and do not directly indicate solve rates on long coding tasks. Official quantized model card

For teams looking to self-host coding agents, this offers an open-weight option they can evaluate directly. The next step is to measure patch success rates, tool-call reliability, and end-to-end task latency on their own repositories. Chinese-language requirement understanding and multi-turn stability after quantization also need independent testing.