JetBrains Releases Mellum2.1 Coding Model with Enhanced Agentic Capabilities

On October 8, JetBrains announced the release of Mellum2.1, a new version of its coding-focused AI model with an emphasis on agentic programming capabilities. The model retains the 12B mixture-of-experts architecture of Mellum2, with 2.5B active parameters, and remains available under the Apache 2.0 license.

The key upgrade in Mellum2.1 is the post-training reinforcement learning stage, which has been expanded from a short final phase into the main training component. Additional training data was added across math, algorithm competitions, science, tool use, and software engineering tasks. JetBrains also built internal reinforcement learning infrastructure, launching millions of sandboxes across thousands of environments during training. The team filtered open datasets beforehand to remove defective tests, unverifiable answers, and misaligned task difficulty.

With the upgrade, Mellum2.1 can explore codebases, edit files, and review its own changes. According to JetBrains, the largest gains are in agentic coding, where the model can identify the root cause of failing tests, draft fixes, and verify the results.

In terms of performance, Mellum2.1 keeps the same architecture and speed as Mellum2, while multi-token prediction (MTP) improves response speed by roughly 1.6x for single requests. In benchmarks against Mellum2, Qwen3.5-9B, and Gemma 4 E4B under the same evaluation setup, Mellum2.1 achieved nearly twice the inference throughput of Qwen3.5-9B under high load.

Mellum2.1 is now available on Hugging Face and can be deployed locally or in private infrastructure, allowing enterprises to keep code and data within their own environment. JetBrains said it will later release GGUF versions for llama.cpp, Ollama, and LM Studio, as well as an MTP speculative decoding component for vLLM.