JetBrains released Mellum2 today, an open-weights 12B Mixture-of-Experts model with 2.5B active parameters per token, available under Apache 2.0 on HuggingFace at JetBrains/mellum-2. Architecture: 28 layers, hidden size 2,304, Grouped-Query Attention (32 query heads, 4 KV heads), Sliding Window Attention on 3 of 4 layers with a 1,024-token window, 131,072 token context, bfloat16 precision, vocabulary 98,304. Activates 8 of 64 experts per token, which is what produces the 2.5B-active number from the 12B total. JetBrains describes the capability set as code generation and editing, debugging, multi-step reasoning, tool use and function calling, agentic coding, and conversational programming assistance, which is a broader scope than Mellum1 (the prior 4B dense completion-focused predecessor).
The framing matters more than any individual benchmark number. JetBrains explicitly positions Mellum2 as a "focal model," a fast, specialized component inside larger AI systems rather than a standalone frontier replacement. Use cases they name: routing/orchestration, RAG summarization, sub-agents in complex workflows, repetitive or latency-sensitive steps that should not be routed through a single large frontier model. The per-token compute is equivalent to a 2.5B dense model, which is the throughput angle the MoE buys you. Benchmark numbers reported: LiveCodeBench v6 37.2, EvalPlus 78.4 (HumanEval+ plus MBPP+), MultiPL-E 67.1, BFCL v3 66.3 (tool use). The comparison cohort in the JetBrains tables is Qwen3.5 4B and 9B, Ministral 3 14B, and OLMo-3 7B, not the Codestral or DeepSeek-Coder or Qwen3-Coder lines, which means head-to-head against the leading code-specialist models is not in the launch material and is the gap you would need to fill on your own evaluation.
Two ecosystem threads. First, the "focal model" framing is the thing to take seriously. The dominant pattern for the last year has been "route every code task through the biggest available frontier model," which is expensive and slow for the 80 percent of code tasks that are not actually hard. The focal-model alternative says: a 2.5B-active, 128K-context specialist that hits 78.4 EvalPlus and 66.3 BFCL handles the repetitive layer cheaply and fast, and you escalate to the frontier only for the hard 20 percent. Mellum2 plus an orchestrator (LangGraph, your own router, or Claude Code's new Dynamic Workflows) is the shape JetBrains is suggesting. Second, Apache 2.0 plus open weights plus a HuggingFace home means commercial deployment, self-hosting, and fine-tuning are all in scope without additional licensing negotiations. For builders who already pay frontier-model rates for code completion or simple refactoring, the question is whether Mellum2 covers enough of your actual code task distribution to justify the routing complexity.
Monday morning, if you're paying frontier-model rates for code completion in IDE plugins or PR bots: deploy Mellum2 in a test path and measure the cost-quality crossover on your real workload. If your distribution skews simple (completion, doc generation, small refactors), the focal-model bet pays off; if it skews complex (full-codebase reasoning, novel debugging), you stay on frontier and don't bother. If you're building a multi-model orchestrator yourself, Mellum2 is a sensible default for the fast specialized slot; the tool-use score (BFCL v3 66.3) suggests it can actually handle function calling, not just code completion. And if you're evaluating open code models, run Mellum2 against your real benchmark before reading the LiveCodeBench-v6 37.2 number, because comparable code-specialist scores from Qwen3-Coder and DeepSeek-Coder lines are not in the JetBrains release material and you need them to make a real decision.