The Pulse
JetBrains Releases Mellum2.1 for Coding Agents
JetBrains has released Mellum2.1, an Apache 2.0 model designed to work in code repositories and support coding agents. In JetBrains’ tests, it resolved 47% of SWE-bench Verified tasks, compared with 2% for Mellum2 and 50% for Qwen3.5-9B.
AI.info Team · Written with AI ·
47% on SWE-bench Verified, against 2% for Mellum2
JetBrains’ new Mellum2.1 model resolves 47% of tasks in its SWE-bench Verified evaluation, up from 2% for its predecessor, Mellum2. The result puts it just below Qwen3.5-9B, which scored 50%, and ahead of Gemma 4 E4B, at 23%, in the company’s comparison.
JetBrains announced the release on October 8, 2026. Mellum2.1 is a 12-billion-parameter mixture-of-experts model with 2.5 billion active parameters, offered under the Apache 2.0 license. Its stated purpose is not just to generate code: JetBrains says it can explore a repository, edit files and check its changes, making it suitable as a worker inside a larger coding-agent system.
Training moved into working repositories
The model’s architecture remains the same as Mellum2. JetBrains says most of the new work went into post-training, especially reinforcement learning: the company ran millions of sandboxed training sessions across thousands of environments and expanded tasks to include software engineering, tool use, math, science and competitive programming.
For software tasks, the model trained in real repositories with shell and file-editing tools. The model card says training rewarded it when tests passed. JetBrains also says it filtered training data to remove tasks with broken tests, unverifiable answers, or difficulty mismatched to the model.
“Mellum2 was fast, but it couldn’t work inside a repository at the level we wanted,” writes Bulat Salimzianov, author of the JetBrains announcement.
Strong gains, but not the top score in every agent test
The SWE-bench Verified result is a large improvement over Mellum2, but JetBrains’ own table shows Mellum2.1 trailing Qwen3.5-9B on two other agentic evaluations. It scored 17.4% on Terminal-Bench 2.1, versus Qwen’s 21.7%, and 28% on SWE-bench Pro, versus 38% for Qwen.
Those figures come from JetBrains’ self-reported evaluation, not an independent audit. The company says it used the same open-source Pi v0.73.1 agent harness, with shell and file tools, for each model. The setup also used each model’s default sampling settings, a 114,000-token context and a limit of 16,000 tokens per turn—details that matter when comparing scores.
Local deployment is available; extra formats are pending
JetBrains says Mellum2.1 is available on Hugging Face for local or self-hosted use. That gives teams a way to run the model on their own infrastructure, rather than sending repository code to an external model service.
JetBrains reports that, under heavy load, Mellum2.1 serves almost twice as many tokens as Qwen3.5-9B; for a single request, it says multi-token prediction makes the model about 1.6 times faster. The company’s post identifies an H200 GPU in its output-speed comparison. GGUF builds for llama.cpp, Ollama and LM Studio, along with the multi-token-prediction head for speculative decoding in vLLM, were listed as coming soon—not yet available in the announcement.
For teams considering it as an agent component, the immediate trade-off is clear in JetBrains’ figures: Mellum2.1 makes a sharp advance over Mellum2 on verified repository fixes, while Qwen3.5-9B still leads it on the two harder agent benchmarks reported. The model weights are available now; the promised GGUF and vLLM acceleration components are not.