JetBrains releases Mellum2.1, an Apache-2.0 12B MoE coding model whose SWE-bench Verified score jumps from 2% to 47% after RL in real repositories
Confirmed
The takeaway
On Oct 8, 2026 JetBrains released Mellum2.1-12B-A2.5B-Thinking, an open-weight (Apache 2.0) mixture-of-experts coding model with 2.5B active parameters and a 131k context.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 10 October 2026
Your AI and this story
- GPT-6 Astra161 days after its cutoff
- Claude Opus 5.5100 days after its cutoff
- Gemini 3.8 Flash191 days after its cutoff
- Grok 4.7130 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 100 days before it.
Key facts
- Architecture: 12B total / 2.5B active MoE, 131,072-token context, Apache 2.0; Hugging Face JetBrains/Mellum2.1-12B-A2.5B-Thinking (+ GGUF repo)
- Model card, Mellum2.1 vs Mellum2: SWE-bench Verified 47.0% vs 2.0%; LiveCodeBench v6 82.0% vs 69.4%; AIME 25/26 83.3% vs 60.1%; GPQA Diamond 64.6% vs 51.0%; BFCL v4 62.3% vs 49.6%; HumanEval+ 91.5% vs 90.9%
- Training: RL ‘at a new scale’ in real environments with shell and file-editing tools, rewarded when tests pass (JetBrains)
- Speed (JetBrains): under heavy load serves almost twice as many tokens as Qwen3.5-9B; about 1.6x faster single requests with multi-token prediction
- Serving: vLLM; GGUF builds for llama.cpp, Ollama and LM Studio
What happened
JetBrains, the maker of IntelliJ and PyCharm, published Mellum2.1, an update of its open Mellum2 coding model. The architecture did not change. The gain comes from post-training: the model practised in real repositories with shell and file-editing tools and was rewarded when tests passed. On SWE-bench Verified it went from almost nothing (2.0%) to 47.0%, by the model card’s numbers.
Why it matters
It is a clear example of how much agentic RL alone can lift a small model. A 2.5B-active model that runs on a laptop now resolves about half of SWE-bench Verified, which makes cheap local sub-agents practical inside IDEs.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 2 primary, 1 press
Primary
- JetBrains AI blog: Mellum2.1 gets to work, a fast open model for coding agentsblog.jetbrains.com, official
- Hugging Face: JetBrains/Mellum2.1-12B-A2.5B-Thinking (model card)huggingface.co, code
Press
- LLM Reference: Mellum2.1 Thinking (release date Oct 8)llmreference.com, press
Changes
- Filed (release date from LLM Reference; the JetBrains blog page shows only “October 2026”)