JetBrains’s Mellum2.1 Turns a Summer of Sandbox RL Into a 47% SWE-bench Small MoE

JetBrains published Mellum2.1 on October 8, 2026, as the next open release of its 12-billion-parameter mixture-of-experts coding model. The architecture is unchanged from Mellum2: 2.5 billion active parameters, Apache 2.0 weights on Hugging Face, and a 131,072-token context. Almost all of the new work sits after pre-training. The company says it spent the summer on reinforcement learning inside real repositories, with millions of sandboxed runs across thousands of environments, so the model can explore a codebase, edit files and check its own changes.

The Hugging Face card for JetBrains/Mellum2.1-12B-A2.5B-Thinking lists 28 layers, a 2,304 hidden size, 64 experts with eight activated, grouped-query attention and a 1,024-token sliding window on three of every four layers. Precision is bfloat16. JetBrains positions the Thinking variant for complex agentic work—repository edits, shell commands and tool calls—as well as hard non-agentic coding, math and reasoning. GGUF builds for llama.cpp, Ollama and LM Studio, plus a multi-token prediction head for speculative decoding in vLLM, are listed as coming soon.

From a short RL stage to the main training loop

In Mellum2, reinforcement learning was a short final stage. In Mellum2.1 it becomes the main post-training phase. JetBrains says it ran many experiments on both methods and data, then kept what held up. New RL tasks cover math, competitive programming, science, tool use and software engineering, mixing open datasets with tasks the company built itself. Open data often arrives with broken tests, unverifiable answers or problems that are too easy or impossible, so every source was filtered before it reached training.

For software engineering, the model trains inside real repositories with a shell and file-editing tools and is rewarded when tests pass. That environment-first recipe is what JetBrains credits for the jump in agentic scores. The company also says post-training left the architecture alone, so serving speed matches Mellum2, and multi-token prediction makes single-request decoding about 1.6 times faster. Under heavy load it reports Mellum2.1 as the fastest model in its comparison group, serving almost twice as many tokens as Qwen3.5-9B.

Shared harness, larger agentic gains

JetBrains evaluated Mellum2.1 Thinking against Mellum2 Thinking, Gemma 4 E4B and Qwen3.5-9B with the same pipeline in thinking mode. All agentic numbers use the open-source Pi v0.73.1 harness with shell and file tools, a 114K-token context and up to 16K tokens per turn. Non-agentic benches use greedy decoding. The company marks every figure as self-reported.

The agentic columns move the most. On SWE-bench Verified, Mellum2.1 reaches 47.0% resolved, against 2.0% for Mellum2, 23.0% for Gemma 4 E4B and 50.0% for Qwen3.5-9B. Terminal-Bench 2.1 rises from 0.6% to 17.4%, still behind Qwen3.5-9B at 21.7%. SWE-bench Pro goes from 0.0% to 28.0%, with Qwen3.5-9B at 38.0%. Coding benches also improve: LiveCodeBench v6 climbs from 69.4% to 82.0%, HumanEval+ to 91.5% and MBPP+ to 79.4%. AIME 25/26, the mean of AIME 2025 and AIME 2026, jumps from 60.1% to 83.3%. Tool-use scores rise on BFCL v4 to 62.3% and on ToolHop to 49.1%. HarmBench falls from 21.5 to 8.5, where lower is better.

Those are JetBrains’s own numbers under one harness, not an independent leaderboard. Qwen3.5-9B still leads on several agentic and knowledge tests, including SWE-bench Verified, Terminal-Bench 2.1, SWE-bench Pro, AIME, GPQA Diamond and MMLU-Redux. Mellum2.1’s pitch is a compact, self-hostable worker that closed most of the gap to that class after a summer of environment RL, not a claim that it tops every table.

Where the company wants it to run

JetBrains frames three uses. First, a capable worker inside larger agent systems that can own pieces of a plan, from finding why a test fails to drafting and checking a fix. Second, a general assistant for everyday questions and hard math. Third, private deployment on local hardware or company infrastructure so code and data stay on-premises. Serving recipes on the model card also show how JetBrains expects the Thinking build to be used day to day: long generations, optional tool calling and a reasoning parser aligned with the Qwen3 family. The company says Mellum2 Thinking was re-evaluated under this same pipeline, so its numbers in the table differ slightly from the earlier Mellum2 technical report.

The release lands in a week already crowded with coding-agent models and harness work. What stands out here is the combination of an unchanged small MoE, an Apache 2.0 license, and a post-training story that is almost entirely about sandboxed repository RL rather than a new architecture. For teams that want an open sub-agent they can serve themselves, Mellum2.1 is the clearest JetBrains checkpoint yet for that job.