Is Mellum2.1 state of the art?Pending benchmarks.Announced Oct 8, 2026; no independent results yet.
JetBrains' open-weights coding model for agents and sub-agents: a 12B mixture-of-experts thinking model with 2.5B parameters active per token and a 131,072-token context. It keeps the Mellum2 architecture unchanged; the update comes from reinforcement-learning post-training in sandboxed coding environments. Weights on Hugging Face under Apache 2.0, with GGUF builds and a multi-token-prediction head for faster vLLM serving. No hosted API is stated; it is meant to run on your own hardware.
The details so far
- Lab
- JetBrains
- Announced
- Oct 8, 2026
- API
- Not yet · None stated; self-hosted (vLLM, llama.cpp, Ollama, LM Studio)
- Open weights
- Published · Hugging Face JetBrains/Mellum2.1-12B-A2.5B-Thinking (BF16, not gated)
- Parameters
- About 12.15B total (12,149,923,072 BF16 parameters), 2.5B active per token
- Type
- Text-only thinking model (reasons in <think> blocks) for coding agents: exploring a repository, editing files and checking its own work
- Architecture
- Mixture of experts, 12B total and 2.5B active: 64 experts with 8 active per token; unchanged from Mellum2
- Context
- 131,072 tokens
- Base model
- JetBrains/Mellum2-12B-A2.5B-Base (per the Hugging Face card); the next version of Mellum2 Thinking, not a fine-tune of it
- Training
- Post-training mostly reinforcement learning, with millions of sandboxed runs across thousands of environments, per JetBrains
- Variants
- JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF (BF16, Q8_0, Q6_K, Q4_K_M and MXFP4_MOE quantizations plus MTP head files) and JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP (the same weights plus a multi-token-prediction head for speculative decoding in vLLM, added Oct 30, 2026)
- License
- Apache 2.0, per the Hugging Face card metadata and text and the JetBrains blog; the Hugging Face repo has no separate LICENSE file
- Related report
- Mellum2 technical report, arXiv 2605.31268
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: in JetBrains' own thinking-mode evaluation, 82.0 on LiveCodeBench v6, 91.5 on HumanEval+, 79.4 on MBPP+, 83.3 on AIME 25/26 and 64.6 on GPQA Diamond.
Lab claim: 47.0 on SWE-bench Verified, 28.0 on SWE-bench Pro and 17.4 on Terminal-Bench 2.1 with the Pi v0.73.1 agent harness, against 2.0, 0.0 and 0.6 for Mellum2 Thinking and 50.0, 38.0 and 21.7 for Qwen3.5 (9B).
Lab claim: 62.3 on BFCL v4, 90.6 on IFEval and 87.8 on MMLU-Redux, with an 8.5% HarmBench harmful rate.
Lab claim: under heavy load it serves almost twice as many tokens as Qwen3.5-9B, and MTP makes a single request about 1.6 times faster.
Sources
LABMellum2.1 Gets to Work: A Fast Open Model for Coding Agents (JetBrains AI blog)blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/ · Oct 8, 2026REPOJetBrains/Mellum2.1-12B-A2.5B-Thinking on Hugging Facehuggingface.co/JetBrains/Mellum2.1-12B-A2.5B-ThinkingREPOJetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF on Hugging Facehuggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUFREPOJetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP on Hugging Facehuggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTPPAPERMellum2 technical report (arXiv)arxiv.org/abs/2605.31268Tracking continues until an independent board lists it.