JetBrains releases Mellum2.1, a fast open model for coding agents

JetBrains released Mellum2.1 on October 8, 2026. The model is the next version of Mellum2, which the company open-sourced in June, and it is built for coding agents. A coding agent is an AI tool that works inside a software project on its own: it explores the files, edits code and checks its own changes. Mellum2.1 is released under the Apache 2.0 license and the model files are openly available to download, so organizations can run it on their own computers or servers.

According to JetBrains' announcement, Mellum2.1 is meant for coding agents and for fast sub-agents that take over parts of a larger agent's plan. The company's example is finding the root cause of a failing test, then drafting a fix and checking it. The model can also act as a general assistant for everyday questions and for math and reasoning problems that need step-by-step work. JetBrains highlights local deployment for cases where code and data must stay fully under the user's control.

Same architecture, training shifted to reinforcement learning

Mellum2.1 keeps the architecture of Mellum2: a mixture-of-experts (MoE) model with 12 billion parameters in total. In this design the model is made of many small sub-networks, or "experts", and only some of them run for each token. According to the model card, 8 of the 64 experts are active at a time and about 2.5 billion parameters are used per step, which makes the model fast for its size. The context window, the amount of text the model can read at once, is 131,072 tokens. Mellum2.1 is a "thinking" model: it writes out its reasoning in a separate block before giving the answer.

According to the company, almost all of the work on this version went into post-training, mainly reinforcement learning (RL). RL went from a short final stage in Mellum2 to the main part of training. Open RL datasets were combined with tasks JetBrains built itself in math, competitive programming, science, tool use and software engineering, and broken tests, unverifiable answers and tasks that were too easy or impossible were filtered out before training. For software engineering, the model trained inside real repositories with a shell and file-editing tools and was rewarded when the tests passed. The announcement says millions of sandboxed runs across thousands of environments were launched during training.

The company's comparisons and speed measurements

JetBrains compared Mellum2.1 with Mellum2 and with two open models of a similar class, Qwen3.5-9B and Gemma 4 E4B, using the same evaluation setup for all of them. According to the company, the biggest gain is in agentic coding: in the model card's table, Mellum2.1 scores 47 percent on SWE-bench Verified, against 2 percent for Mellum2. On the same test Qwen3.5-9B leads with 50 percent. On the coding test LiveCodeBench v6, Mellum2.1 has the highest score in the table, while Qwen3.5-9B scores higher on several math, knowledge and agentic tests.

Because the architecture did not change, Mellum2.1 runs as fast as Mellum2. In JetBrains' measurements, Mellum2.1 is the fastest model in the group under heavy load and serves almost twice as many tokens as Qwen3.5-9B. The company says multi-token prediction (MTP) makes a single request about 1.6 times faster; the MTP head needed for this speedup in vLLM is listed as coming soon.

Options for running it locally

The model is on Hugging Face as JetBrains/Mellum2.1-12B-A2.5B-Thinking. The model card includes commands for serving it with vLLM, with a separate command for those who want tool calling. A separate repository holds GGUF files for llama.cpp, Ollama, LM Studio and similar tools. These come in several quantization levels; the 4-bit Q4_K_M file that JetBrains recommends is about 8 GB. Ollama users can start the model with one command: ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M. The blog post says GGUF builds are coming soon, while the GGUF repository on Hugging Face is already available.

All the published comparisons were run by JetBrains with a single setup. The model card lists English as the only language, so it is worth testing the model on your own examples before using it with other languages. The announcement does not say which JetBrains products will use the model. As a thinking model it also produces reasoning text ahead of the answer, and the example code in the model card sets a high output limit (81,920 tokens) for this. For a fuller picture, follow evaluations from independent researchers as well.

Sources: JetBrains: Mellum2.1 Gets to Work, a fast open model for coding agents, Hugging Face: Mellum2.1 Thinking model card, Hugging Face: Mellum2.1 Thinking GGUF.