JetBrains Open-Sources Mellum2 to Go Where Claude Code Can’t

Why It Matters

Mellum2 provides a high‑performance, self‑hosted alternative for enterprise AI tooling, reducing reliance on external APIs and lowering inference costs at scale.

Key Takeaways

• Mellum2 open‑sourced with 12B parameters, 2.5B active per token
• MoE architecture yields 21% faster concurrent inference than Qwen2.5-7B
• Offers “instruct” and “thinking” variants for direct answers and reasoning traces
• Enables on‑prem deployment, avoiding reliance on Claude Code’s cloud APIs

Pulse Analysis

JetBrains' decision to open‑source Mellum2 reflects a broader shift toward specialized, focal models that complement frontier LLMs. By leveraging a Mixture‑of‑Experts (MoE) architecture, Mellum2 maintains the capacity of a 12‑billion‑parameter model while activating only 2.5 billion parameters per token, delivering near‑dense‑model performance with substantially lower compute overhead. This design choice translates into tangible speed gains: on a single NVIDIA H100 GPU, Mellum2 matches Qwen2.5‑7B in single‑request throughput and surpasses it by 21% under realistic concurrent workloads, making it attractive for high‑frequency code‑completion and agentic pipelines.

Beyond raw speed, Mellum2's dual variants—"instruct" for straightforward answers and "thinking" for step‑by‑step reasoning—address the nuanced needs of modern development teams. The "thinking" version achieved a 78.4% score on the EvalPlus benchmark, outpacing comparable models such as Qwen3.5‑9B and Seed‑Coder‑8B. While its narrower training focus yields lower performance on broad knowledge tasks, this trade‑off is intentional, optimizing the model for software‑engineering contexts where code generation and documentation retrieval dominate. Enterprises can thus deploy a model that excels where they need it most without the latency and cost penalties of larger, general‑purpose LLMs.

The strategic importance of open‑sourcing cannot be overstated. Unlike Claude Code or OpenAI's Codex, which require API calls to external services, Mellum2's Apache 2.0 license permits on‑premises deployment, granting organizations full control over data privacy, compliance, and operational costs. As AI becomes embedded deeper into CI/CD pipelines and internal tooling, the ability to host a performant, code‑focused model internally could become a differentiator for firms seeking to scale AI‑driven development while mitigating vendor lock‑in. JetBrains' move positions it as a key enabler for the next generation of autonomous software engineering stacks.