JetBrains Opens Mellum2 Thinking, a 12B Coding Model Built for Agentic AI Speed

JetBrains has open-sourced Mellum2 Thinking, a 12-billion-parameter Mixture-of-Experts model that activates only 2.5B parameters per token and emits explicit chain-of-thought traces before answering. It is the reasoning sibling in a six-checkpoint family aimed squarely at the plumbing of agentic AI systems: routing, retrieval, sub-agents, and code-heavy workloads that have to run on your own hardware.

The pitch is unusual. Rather than chase frontier benchmarks, JetBrains is positioning Mellum2 as what they call a focal model, optimized for the high-frequency, latency-sensitive calls that pile up around a larger reasoning model. A single frontier model can be powerful, but production systems often need several specialized components working together: retrievers, routers, code-aware models, validators, tool callers, and larger reasoning models, and JetBrains thinks of Mellum2 as a focal model: a fast, well-scoped model optimized for high-frequency tasks inside larger AI systems.

What shipped

JetBrains published the Mellum 2 family on Hugging Face as a Mixture-of-Experts coding model at 12B total parameters with 2.5B active per token, released under Apache 2.0, with six variants shipping simultaneously: Base, Base-Pretrain, Instruct, Instruct-SFT, Thinking, and Thinking-SFT. The Thinking variant is the one most teams will reach for when they want the model to deliberate.

The headline architecture numbers:

• A Mixture-of-Experts architecture with 64 experts that activates 8 experts per token, combined with sliding-window and full attention layers, and a context length of 131,072 tokens.