Mellum 2.1
JetBrains' 12B open MoE coding model, RL-trained in real repos to explore codebases, edit files, and check its own work.
About
Mellum 2.1 is JetBrains' open-weight coding model announced October 8, 2026, the successor to the Mellum 2 it open-sourced that June. The architecture is unchanged: a 12B mixture-of-experts with 2.5B active parameters (64 experts, 8 active), a 131,072-token context, and a thinking variant that emits its chain of thought before answering. Everything that changed is post-training: a summer of reinforcement learning across millions of sandboxed runs in real software environments, aimed at repo work rather than autocomplete.
JetBrains reports SWE-bench Verified jumping from 2.0% (Mellum 2 Thinking) to 47.0% (Mellum 2.1 Thinking) in its own matched evaluation, plus 82.0 on LiveCodeBench v6. Independent coverage notes the honest framing: JetBrains benchmarks against its own weight class (Qwen3.5-9B, Gemma 4) rather than frontier giants, and Qwen3.5-9B still leads on the hardest agentic tasks. Weights are Apache 2.0 on Hugging Face with official GGUF builds from 7.0GB, served via vLLM, SGLang, Ollama, or LM Studio.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SWE-bench Verified: 47.0% (JetBrains-reported, up from 2.0% for Mellum 2 Thinking)
- LiveCodeBench v6: 82.0 (JetBrains-reported, via MarkTechPost summary)
- Terminal-Bench 2.1: 17.4 (JetBrains-reported; Qwen3.5-9B scored 21.7 on the same table)
- All figures are vendor-reported from JetBrains' own evaluation
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: real JetBrains release Oct 8, 2026 with Apache-2.0 weights and official GGUF builds on Hugging Face; detailed launch post describing the RL post-training methodology; benchmark table with matched baselines. Open weights. Launch facts corroborated across the JetBrains blog, MarkTechPost, and independent coverage. Strengths: honest same-weight-class benchmarking, strong packaging for local deployment, credible vendor. Docked for: (1) all benchmark figures are vendor-reported with no independent replication yet; (2) still trails Qwen3.5-9B on hard agentic software tasks in JetBrains' own table.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · JetBrains release 2026-10-08; 12B MoE, 2.5B active, 131K context
- · Apache-2.0 weights + official GGUF builds; vLLM/SGLang/Ollama support
- · JetBrains-reported: SWE-bench Verified 47.0 (from 2.0), LiveCodeBench v6 82.0
- · All benchmarks vendor-reported; just released so adoption is early