JetBrains’ Open‑Source 12B Code Model Mellum2: Private Deployment and High‑Throughput Over Claude Code

JetBrains announced the open‑source release of Mellum2, a 12‑billion‑parameter code model that succeeds the earlier 4‑billion‑parameter Mellum. Designed for AI‑agent infrastructure, Mellum2 targets private, on‑premise deployment and aims to address limitations of Claude Code and similar services.

Designed for Large‑Scale Speed

Mellum2 uses a Mixture‑of‑Experts (MoE) architecture with 64 expert sub‑models. Each token activates only 2.5 b parameters, routing the token to a subset of experts, which preserves the model’s overall capacity while significantly improving inference speed.

"The design of Mellum2 is intended to meet the broader set of tasks engineering teams face when deploying AI, such as model coordination, sub‑agent workloads, and context compression in retrieval pipelines."

In benchmarks on a single H100 GPU, Mellum2 processes 192 tokens per second, virtually matching Qwen2.5‑7B’s 193 t/s. Under realistic concurrent loads, Mellum2 outperforms Qwen2.5‑7B by 21 % and Qwen3‑8B by 79 %. Because only 2.5 b parameters are active per token, the cost profile resembles that of a dense 2.5 b model rather than a full 12 b model, which is crucial for high‑volume routing in large‑scale agent systems.

For functional code generation, Mellum2’s “thinking” variant scores 78.4 % on the EvalPlus suite (HumanEval+ + MBPP+), surpassing Qwen3.5‑9B (71.8 %) and the code‑focused Seed‑Coder‑8B (73.8 %). However, on broader reasoning and knowledge benchmarks such as GPQA Diamond and MMLU‑Redux, larger models like Qwen3.5‑9B retain an advantage, a trade‑off the authors acknowledge.

The authors describe Mellum2 as a “focus model”: fast and specialized rather than broadly competitive with frontier large models. Two post‑training variants are provided—a direct‑answer “instruction” version and a chain‑of‑thought “thinking” version that generates explicit reasoning traces for multi‑step tasks.

Mellum2 is released under the Apache 2.0 license and hosted on Hugging Face, including the full technical report, architecture decisions, and training details. This open‑source stance gives enterprises the option to self‑host the model and control the AI stack, though its adoption will depend on the demand for self‑managed AI infrastructure in software engineering workflows.