JetBrains2026-10-08 16:49:46JetBrains open-sources Mellum2.1 coding agent model, topping Qwen3.5-9B on several benchmarksJetBrains has released Mellum2.1, an open-source model built for coding agents and fast subagents. The company describes it as a 12 billion-parameter mixture-of-experts reasoning model with 2.5 billion active parameters per token. According to the announcement, Mellum2.1 outperformed its predecessor Mellum2 in 15 of 17 benchmarks and ranked ahead of Qwen3.5-9B on several coding evaluations, including LiveCodeBench v6. JetBrains published the model on Hugging Face under the Apache 2.0 license and said the architecture remains consistent with Mellum2. The main upgrade came from reinforcement learning training conducted in real software environments. JetBrains said the model is small enough to be self-hosted and can explore code repositories, edit files, and inspect its own changes. The company added that users can run it on their own GPUs through vLLM or SGLang. Its speed tests were based on a single NVIDIA H200. Mellum2.1 supports a 131,072-token context window and is aimed at agentic work, general reasoning assistants, and private self-hosted deployments, according to MarkTechPost as cited by Techub.10
JetBrains2026-10-09 08:06:01JetBrains open-sources Mellum2.1 after posting a higher LiveCodeBench score than Qwen3.5-9BJetBrains has released Mellum2.1, an open-source coding model positioned as an upgrade to Mellum2, which the company introduced in June this year. The new version keeps the same 12 billion total parameters and activates 2.5 billion parameters per generation, but JetBrains said it performs better at identifying code issues, editing files, and checking results. The company said it did not change the model architecture and instead improved performance mainly through reinforcement learning. According to JetBrains, the team ran millions of sandbox tasks across thousands of environments during training, with the model using terminal and file-editing tools inside real code repositories and receiving rewards when tests passed. In JetBrains’ own benchmarks, Mellum2.1 raised its success rate on SWE-bench Verified from 2% to 47%, slightly below Qwen3.5-9B’s 50%. On LiveCodeBench v6, however, Mellum2.1 scored 82%, ahead of Qwen3.5-9B’s 75.4%. JetBrains also said that under heavy-load inference on a single H200, the new model’s output throughput was close to twice that of Qwen3.5-9B. Model weights and a GGUF quantized version are now available on Hugging Face under an Apache 2.0 license for local deployment.20
