Mellum 2.1
Released on October 8, 2026
Changes from previous version
Post-training is the change, mostly reinforcement learning inside real repositories. On the shared JetBrains harness, SWE-bench Verified moves from 2.0% to 47.0%, Terminal-Bench 2.1 from 0.6% to 17.4%, and SWE-bench Pro from 0.0% to 28.0%. With multi-token prediction, one request on one H200 (FP8, 2,304 input tokens and 256 output tokens) reaches 557 output tokens per second, about 1.6 times the 339 tokens per second without it.
Release Summary
Open-weight thinking checkpoint for self-hosted coding agents. Same 12B mixture-of-experts architecture as Mellum 2, with 2.5B active parameters, Apache 2.0 weights, and a 131,072-token context. JetBrains reports 47.0% on SWE-bench Verified and 17.4% on Terminal-Bench 2.1, up from 2.0% and 0.6% for Mellum 2 Thinking on the same Pi v0.73.1 harness. LiveCodeBench v6 is 82.0. Qwen3.5-9B still leads that table on SWE-bench Verified at 50.0 and SWE-bench Pro at 38.0. Weights: JetBrains/Mellum2.1-12B-A2.5B-Thinking.
Timeline
Mellum 2.1 released
JetBrains publishes Mellum 2.1 Thinking weights on Hugging Face under Apache 2.0 for local coding agents and sub-agents.