JetBrains open-sources Mellum 2, featuring 12B total parameters

JetBrains has announced that it is open-sourcing its new machine learning model designed for software engineering systems, Mellum2.

This comes a little over a year after the company open-sourced the original Mellum in 2025, a relatively small model with 4 billion parameters. Mellum2, its successor, comes with 12B parameters in total, though JetBrains promises that it keeps computation fast by utilizing only 2.5B active parameters per token. The engineering team deployed a sparse Mixture-of-Experts framework to achieve these fast inference speeds on standard hardware.

Unlike Mellum, which operated as a "focal" model (focal, as in it only focused on a single task like code completion inside an editor), Mellum2 functions as a full-fledged coding assistant that can generate and edit code, call external tools, execute multi-step agentic workflows, hold long conversations, and use explicit reasoning. Developers can deploy the base, instruct, and thinking versions under the permissive Apache 2.0 license to retain full data control. Here's a quick spec sheet of how Mellum2 compares to Mellum:

JetBrains notes that Mellum2 is great for things like routing and orchestrating AI workloads, building low-latency RAG pipelines, powering fast sub-agents in complex workflows, and enabling private, local AI deployment. The team optimized the pre-training stage with a three-phase data curriculum that progressively shifted the mixture from diverse web data toward curated code and mathematical content.

In other JetBrains news, JetBrains is discontinuing DataSpell, the popular IDE that Python developers use. JetBrains decided to sunset DataSpell on May 28, to consolidate all data science features directly into PyCharm Pro. The company will automatically convert active subscriptions on September 1, 2026.