JetBrains drops Mellum2, a 12B MoE model built for speed

Mellum2 is an open-source 12B Mixture-of-Experts model from JetBrains, built from the ground up for low-latency text and code workloads.

Instead of chasing frontier-scale benchmarks on everything, Mellum2 is laser-focused on the glue work inside software engineering stacks: prompt routing, context compression for RAG, sub-agent validation, and high-throughput coding tasks.

How it gets so fast

The secret sauce is its sparse Mixture-of-Experts setup. While the total model footprint sits at 12B parameters, Mellum2 activates only 2.5B parameters per token during inference.

That gives you the capacity of a 12B model with the compute footprint of a lightweight 2.5B model. According to JetBrains, Mellum2 runs more than 2x faster than similarly sized open models while keeping competitive scores across code generation, reasoning, science, and math benchmarks.

JetBrains also skipped multimodal data entirely. The model was trained from scratch strictly on natural language and code, keeping the architecture compact and dedicated to engineering jobs.

Where it fits into your stack

Modern AI systems rarely rely on just one model anymore. Mellum2 is designed to be the workhorse for high-frequency plumbing tasks.

For orchestration and routing, it can classify incoming prompts, pick relevant tools, and manage execution flows without burning expensive API calls on frontier models.

For RAG pipelines, it handles context compression, rapid summarization, and retrieval post-processing where response latency is critical.

For autonomous workflows, Mellum2 can act as a sub-agent handling intermediate steps like planning, syntax validation, and context restructuring.

Because it ships under the Apache 2.0 license, you can also self-host it on your own infrastructure to keep proprietary codebases completely private.

How to run it

The model weights are ready to download right now on Hugging Face under the JetBrains collection.

For teams that want to check out the underlying architecture, training setup, and full evaluation numbers, JetBrains also published the complete technical report on arXiv (2605.31268).

Why it matters

We think Mellum2 is a smart, pragmatic release. Production software engineering systems cannot tolerate multi-second latency for simple intermediate steps, nor does it make financial sense to invoke giant frontier models just to route a prompt.

If you are building code agents, IDE integrations, or internal RAG pipelines, Mellum2 gives you an open, fast component that slots directly into your multi-model architecture to cut latency and slash serving costs.