JetBrains open-sources Mellum2.1: surpasses Qwen3.5-9B on LiveCodeBench test.

Beating AI News Flash: JetBrains releases open-source coding model Mellum2.1. Compared with Mellum2 launched in June this year, the new version still has 12 billion total parameters and activates only 2.5 billion per generation, but is better at finding code issues, modifying files, and checking results.

JetBrains did not change the model architecture and mainly improved capabilities through reinforcement learning. The team ran millions of sandbox tasks across thousands of environments. During training, the model called terminal and file editing tools in real code repositories and received rewards for successfully passing tests.

JetBrains' internal testing shows that the success rate on SWE-bench Verified (a real software issue fixing test) rose from 2% in the old version to 47%, slightly below Qwen3.5-9B's 50%. However, on the LiveCodeBench v6 coding test, Mellum2.1 scored 82%, exceeding Qwen's 75.4%. JetBrains also claims that when running under high load on a single H200, the new model's output throughput is nearly twice that of Qwen3.5-9B.

The model weights and GGUF quantized version have been open-sourced on Hugging Face under the Apache 2.0 license, allowing developers to deploy locally.