JetBrains ships Mellum 2.1, a reasoning MoE for code

JetBrains has released Mellum 2.1, the latest version of its code-focused language model line. The new build is a mixture-of-experts design with 12B total parameters but only 2.5B active per token, and it adds a reasoning-oriented "Thinking" variant aimed at harder programming tasks.

The sparse layout is the headline here. By routing each token through a small fraction of the network, an MoE model can offer the capacity of a larger system while keeping inference costs closer to a 2.5B dense model. For a company whose core business is developer tooling, that efficiency matters: code assistance is latency-sensitive and often runs at scale across IDEs.

Why it matters

• Permissive licensing. Mellum 2.1 ships under Apache-2.0, making it straightforward to fine-tune, self-host, or embed in commercial products.
• Reasoning for code. The Thinking configuration signals a push toward multi-step problem solving rather than pure autocompletion.
• Lean active footprint. At 2.5B active parameters, it is designed to be practical to run outside of heavyweight datacenter setups.

JetBrains has positioned Mellum as a purpose-built family rather than a general-purpose chatbot, and this update continues that focus on software development workloads. Teams evaluating open models for coding can find the weights and details on the Hugging Face repository.