JetBrains releases Mellum2.1

Mellum2.1 is designed to explore code repositories, edit files and check its own changes, according to JetBrains[1]. It has 12 billion total parameters and uses a mixture-of-experts architecture, with 2.5 billion parameters activated per token[3]. JetBrains says the architecture is unchanged from Mellum2, so the release focuses on changes to training rather than a redesigned architecture[1]. The model is released under the Apache 2.0 license[1].

JetBrains says reinforcement learning became the main part of Mellum2.1’s training, replacing its earlier role as a short final stage[1]. The new tasks span math, competitive programming, science, tool use and software engineering[1]. For software-engineering training, Mellum2.1 worked in real repositories using shell and file-editing tools, and received rewards when tests passed[3]. JetBrains says training involved millions of sandbox runs across thousands of environments[1]. These details describe the training setup; they do not by themselves establish how the model will perform on a particular team’s codebase.

MarkTechPost reports that Mellum2.1 beat Mellum2 on 15 of the 17 benchmarks it lists, and notes that the reported benchmark results were self-reported by JetBrains[3]. JetBrains’ SWE-bench Verified score rose from 2.0 to 47.0, while its SWE-bench Pro score rose from 0.0 to 28.0[3]. Qwen3.5-9B scored 50.0 on SWE-bench Verified, above Mellum2.1[3]. Mellum2.1 scored 82.0 on LiveCodeBench v6, ahead of Qwen3.5-9B and Gemma 4 E4B in the reported comparison[3]. Its Terminal-Bench 2.1 score, however, was below Qwen3.5-9B’s[3].

Mellum2.1 is available on Hugging Face, and JetBrains says users can run it locally or on their own infrastructure to keep code and data under their control[1]. MarkTechPost says users can serve the model on their own GPUs using vLLM or SGLang[3]. JetBrains said GGUF builds for llama.cpp, Ollama and LM Studio were coming soon, rather than stating that those builds were already available[1]. JetBrains also says multi-token prediction makes the model about 1.6 times faster for a single request, and that under heavy load it serves almost twice as many tokens as Qwen3.5-9B[1].