JetBrains releases Mellum2.1, a coding AI model focused on agentic programming
IT之家 reports that JetBrains announced the programming AI model Mellum2.1 in a blog post on October 8, with a focus on strengthening agentic coding. As a developer, I have been following open models like this that can run locally, so I am summarizing the key points here.
The model keeps the Mellum2 architecture: a 12B mixture-of-experts (MoE) model with 2.5B active parameters, still released under the Apache 2.0 license. The main change in training is that the reinforcement learning stage after pretraining has been expanded from a short finishing step into a core part of training. JetBrains added data for math, algorithmic competitions, science, tool use, and software engineering, and built internal RL environments, launching millions of sandboxes across thousands of environments during training. Before training, tasks with flawed tests, unverifiable answers, or unsuitable difficulty are filtered out.
According to the official description, it can explore codebases, edit files, and check its own changes. The biggest improvement is in agentic coding: it can identify the root cause of failing tests, draft a fix, and verify the result. On speed, the architecture is unchanged, but multi-token prediction (MTP) makes single-request generation about 1.6 times faster. In comparisons with Mellum2, Qwen3.5-9B, and Gemma 4 E4B under the same evaluation settings, Mellum2.1's output throughput on a single H200 under high load is close to twice that of Qwen3.5-9B.
The model is on Hugging Face and can be deployed locally or on your own infrastructure, so companies can keep their code and data in their own environment. JetBrains plans to release GGUF versions for llama.cpp, Ollama, and LM Studio, along with an MTP speculative decoding component for vLLM.
I am looking forward to the GGUF version, since ordinary developers could then try it on their own machines. When you choose a coding model, which matters more to you: accuracy or deployment cost?
Source: IT之家
The model keeps the Mellum2 architecture: a 12B mixture-of-experts (MoE) model with 2.5B active parameters, still released under the Apache 2.0 license. The main change in training is that the reinforcement learning stage after pretraining has been expanded from a short finishing step into a core part of training. JetBrains added data for math, algorithmic competitions, science, tool use, and software engineering, and built internal RL environments, launching millions of sandboxes across thousands of environments during training. Before training, tasks with flawed tests, unverifiable answers, or unsuitable difficulty are filtered out.
According to the official description, it can explore codebases, edit files, and check its own changes. The biggest improvement is in agentic coding: it can identify the root cause of failing tests, draft a fix, and verify the result. On speed, the architecture is unchanged, but multi-token prediction (MTP) makes single-request generation about 1.6 times faster. In comparisons with Mellum2, Qwen3.5-9B, and Gemma 4 E4B under the same evaluation settings, Mellum2.1's output throughput on a single H200 under high load is close to twice that of Qwen3.5-9B.
The model is on Hugging Face and can be deployed locally or on your own infrastructure, so companies can keep their code and data in their own environment. JetBrains plans to release GGUF versions for llama.cpp, Ollama, and LM Studio, along with an MTP speculative decoding component for vLLM.
I am looking forward to the GGUF version, since ordinary developers could then try it on their own machines. When you choose a coding model, which matters more to you: accuracy or deployment cost?
Source: IT之家
Comments 0
Leave the first comment, in any language.
Log in to comment