JetBrains Released Mellum2.1 Open-Source Model
The 12B parameter model introduces specialized software environment training for coding agents.
Updated on Oct. 9, 2026 in Artificial Intelligence
Live Poll
Would you use self-hosted AI models to assist with your own coding or technical projects?
JetBrains has released Mellum2.1, an open-source mixture-of-experts model optimized for reasoning and coding agent workflows. The model is currently available for deployment via standard inference software.
Why it matters
Mellum2.1 improves performance in software-heavy tasks by utilizing reinforcement learning specifically tailored for coding environments. Its architecture balances a large 12B parameter total with a leaner activation footprint to improve efficiency in agentic applications.
The 12B parameter model utilizes a mixture-of-experts architecture that activates 2.5B parameters per token using a router mechanism. It supports a 131,072-token context window and yields a score of 82.0 on LiveCodeBench v6 and 47.0 on SWE-bench Verified.
The players
JetBrains
A software company focused on integrated development environments and coding tools that is now expanding into generative AI research.
The details
Mellum2.1 employs a mixture-of-experts — a design where only a subset of neural network components is activated for any given input — to achieve high reasoning capability with lower compute requirements. A router mechanism selects 8 out of 64 total experts for each token processed, enabling the model to scale its knowledge without proportional increases in latency. The model is compatible with inference engines like vLLM and SGLang, and it outputs a chain of thought — a sequence of logical steps leading to a final conclusion — before presenting its final response.
Timeline
JetBrains open-sourced the original Mellum2 in June 2026.
Mellum2.1 was released on October 8, 2026.
The Tech Race
Mellum2.1 enters a competitive market where small, expert-based models are increasingly benchmarked against coding-specific tasks. JetBrains is positioning this release to compete with specialized reasoning assistants by emphasizing its performance in simulated software environments.
Developers can integrate the model today using existing inference engines like vLLM and SGLang, with minimal GGUF builds starting at 7.0 GB. The model is released under an Apache 2.0 license, allowing for wide implementation in open-source coding agents.
The takeaway
JetBrains has optimized Mellum2.1 for the specific reasoning patterns required by coding agents rather than general-purpose chat. Watch for the forthcoming release of the multi-token prediction head to see how the model's inference throughput changes in production environments.
What happens next
JetBrains plans to release a multi-token prediction head for vLLM speculative decoding, which will further optimize inference speeds.
Further reading
For broader trends in model development, visit the Artificial Intelligence section.
Live Poll
Would you use self-hosted AI models to assist with your own coding or technical projects?
