Artificial intelligence news.
Straight from the original announcements.

JetBrains releases Mellum2.1, an open 12B model built for coding agents

October 8, 2026, 14:00By the AI News Bites Editorial TeamSource: JetBrains AI
JetBrains releases Mellum2.1, an open 12B model built for coding agents
Image: JetBrains AI

JetBrains has released Mellum2.1, an updated version of its open-source 12B mixture-of-experts model, aimed at coding agents and fast sub-agents that run on users' own hardware. According to the company, the architecture is unchanged from Mellum2, but a summer of reinforcement learning in real environments now lets the model explore a codebase, edit files and check its own changes.

Key points
  • Mellum2.1 is a post-training update of JetBrains' open 12B mixture-of-experts model with 2.5B active parameters, under Apache 2.0.
  • Reinforcement learning became the main part of training, with millions of sandboxed runs across thousands of environments, according to JetBrains.
  • JetBrains says the biggest gain over Mellum2 is in agentic coding, compared against Qwen3.5-9B and Gemma 4 E4B.
  • Under heavy load it serves almost twice as many tokens as Qwen3.5-9B, and MTP makes single requests about 1.6 times faster, per the company.
  • The model is on Hugging Face; GGUF builds and the vLLM MTP head are coming soon.

Same architecture, very different training

Mellum2.1 follows the model JetBrains open-sourced in June. The architecture has not changed since version 2: it is a compact model with 12B parameters, 2.5B of them active, released under the Apache 2.0 license.

What changed is everything after pre-training. JetBrains says Mellum2 was fast but could not work inside a repository at the level the team wanted. After what it describes as a summer of reinforcement learning (RL) in real environments, with millions of sandboxed runs across thousands of environments, the company says Mellum2.1 can explore a codebase, edit files and verify its own changes.

What went into the RL work

According to JetBrains, almost all of the effort for this version went into post-training. The company lists three main changes:

  • RL at a new scale: reinforcement learning went from a short final stage to the main part of training. The team ran many experiments on both methods and data and kept what held up.
  • More data, filtered harder: new RL tasks were added in math, competitive programming, science, tool use and software engineering, combining open RL datasets with tasks JetBrains built itself. The company notes that open data often has broken tests, unverifiable answers, or tasks that are too easy or impossible, so every source was filtered before training.
  • Real environments: JetBrains built infrastructure to run thousands of RL environments in-house and launched millions of sandboxes during training.

Performance and speed

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. The post does not give benchmark figures. The company says the biggest gain over Mellum2 is in agentic coding, with improvements also in coding, competitive programming, math, tool calling and general knowledge.

On speed, post-training did not touch the architecture, so Mellum2.1 is as fast as Mellum2, and multi-token prediction (MTP) makes it faster. JetBrains says that under heavy load Mellum2.1 is the fastest model in the group and serves almost twice as many tokens as Qwen3.5-9B. For a single request, MTP makes it about 1.6 times faster, according to the company.

Use cases and availability

JetBrains positions Mellum2.1 as a capable worker inside agentic systems, handling parts of an agent's plan from identifying the root cause of a failing test to drafting and checking a fix. It is also described as a general assistant that handles everyday questions and works through hard math and reasoning problems step by step. For teams that want control over code and data, it can be run locally or on their own infrastructure.

Mellum2.1 is available on Hugging Face. GGUF builds for llama.cpp, Ollama and LM Studio, as well as the MTP head for speculative decoding in vLLM, are listed as coming soon. JetBrains invites developers building coding agents, sub-agents or AI tools on their own infrastructure to try the model and send feedback, which it says will shape the next version.

Related news

AWS shows how to route coding agent requests with NeMo Switchyard
For developersNovember 13, 2026, 23:30

AWS shows how to route coding agent requests with NeMo Switchyard

AWS has published a sample deployment that routes coding agent requests between models on Amazon Bedrock using NVIDIA NeMo Switchyard and LiteLLM. The approach measures…

Block brings people and AI agents together with Buzz and Unity Gateway
Products and toolsNovember 12, 2026, 15:59

Block brings people and AI agents together with Buzz and Unity Gateway

Block is using its open source Buzz workspace and Databricks Unity Gateway to let employees collaborate with AI agents, colleagues, and multiple AI models in one…

JetBrains Brings Air Agent Workspace to Its IDEs
Products and toolsNovember 11, 2026, 10:38

JetBrains Brings Air Agent Workspace to Its IDEs

JetBrains has made its Air agent workspace available in JetBrains IDEs with the 2026.3 release. Air lets developers run supported coding agents inside the IDE, track…