JetBrains Open-Sources Mellum2: 12B MoE at 2x Speed for Agentic Sub-Task Routing
JetBrains released Mellum2 on June 1 under Apache 2.0. The model targets the infrastructure layer of agentic AI systems: the calls that happen between the big expensive reasoning model and the user. Routing decisions, document chunking, sub-agent coordination, retrieval scoring, and private on-premises deployments where Claude Code or GPT-5.5 cannot go.
The model is a 12B Mixture-of-Experts with 2.5B active parameters per token — purpose-built for high-throughput, latency-sensitive workloads. JetBrains reports 2x faster inference than comparable open models in the 4B–14B range, with a 21% throughput advantage over Qwen2.5-7B and a 79% advantage over Qwen3-8B under concurrent load on a single H100.
Benchmark Results (Self-Reported)
All numbers below are reported by JetBrains in their technical report. Comparison set is open-weight models in the 4B–14B range.
| Benchmark | Mellum2 | Qwen3.5 9B | Ministral 3 14B | OLMo-3 7B |
|---|---|---|---|---|
| LiveCodeBench v6 | 37.2% | 63.7% | 42.4% | 28.2% |
| EvalPlus (HumanEval+ / MBPP+) | 78.4% | 71.8% | 74.1% | 67.3% |
| MultiPL-E | 67.1% | 67.1% | 71.5% | 36.1% |
| BFCL v3 (tool use) | 66.3% | 70.5% | 52.7% | 41.9% |
| BFCL v4 (tool use, extended) | 44.2% | 60.6% | 38.8% | 19.8% |
| AIME 2025+2026 | 41.7% | 58.3% | 33.3% | 40.0% |
| IFEval | 75.8% | 83.9% | 67.3% | 83.2% |
LiveCodeBench is the notable weak spot — Mellum2 trails Qwen3.5 9B by 26 points on competitive coding problems. EvalPlus tells a different story: 78.4% against Qwen3.5 9B’s 71.8%, suggesting Mellum2 is stronger on practical code generation than on competitive algorithmic tasks. Tool use via BFCL v3 at 66.3% is competitive for its size class. BFCL v4 — which adds web search and memory subtasks — drops more sharply, indicating Mellum2 is stronger on structured function calling than on open-ended tool chains.
Design Brief: Agentic Infrastructure
The original Mellum was a 4B dense code completion model shipped inside JetBrains IDEs. Mellum2 is a different product: a general-purpose model optimised for the latency-sensitive, high-throughput calls that make up the majority of tokens in a production agentic pipeline. Most agent architectures invoke a model dozens of times per task for classification, summarisation, retrieval reranking, and tool selection — none of which require a 70B reasoning model.
The MoE architecture keeps active parameters low (2.5B per token) while total model capacity (12B) enables broader knowledge coverage than a dense 2.5B model would have. The inference speed advantage comes from smaller active footprint plus JetBrains’ claim that the model was explicitly trained and optimised for concurrent production workloads rather than benchmark maximisation.
Availability
Weights: Hugging Face (JetBrains/mellum-2 collection), base, instruct, and thinking checkpoints.
License: Apache 2.0 — commercial use, modification, and redistribution permitted.
Technical report: full architecture and training pipeline documentation released alongside the weights.
Mellum2 occupies a specific slot in the agentic stack — it is not competing with Claude Opus 4.8 or GPT-5.5 for reasoning tasks. It is competing with Qwen3.5 4B, Gemma 4 2B, and similar models for the ambient inference workload that surrounds the frontier call. At 2x the throughput and competitive quality on code generation and tool use, it makes a credible case for that workload on private infrastructure where closed-model APIs are off the table.