The launch
On October 8, 2026, JetBrains, maker of IntelliJ IDEA, PyCharm and the Kotlin language, released Mellum2.1, the new version of its open coding model. It's a 12-billion-parameter Mixture-of-Experts with just 2.5 billion active per token, under Apache 2.0 (commercial use allowed).
The architecture is the same as Mellum2, opened in June. What changed is post-training. According to JetBrains, Mellum2 was fast but couldn't work inside a repository at the level it wanted. After a summer of reinforcement learning (RL) in real environments, Mellum2.1 can explore a codebase, edit files and check its own changes.
What changed
- RL as the main part of training, no longer a short final stage.
- More data, filtered harder: new tasks in math, competitive programming, science, tool use and software engineering. JetBrains says open data often has broken tests, unverifiable answers or tasks that are too easy, so everything was filtered.
- Real environments at scale: in-house infrastructure running thousands of RL environments, with millions of sandboxes over training.
Performance and speed
JetBrains compared Mellum2.1 with Mellum2, Qwen3.5-9B and Gemma 4 E4B using the same evaluation setup. The biggest gain was in agentic coding, with improvements in coding, competitive programming, math, tool calling and general knowledge. Exact numbers appear in charts on the blog; they are the company's own results.
On speed:
- Under heavy load, it was the fastest in the group, serving almost twice as many tokens as Qwen3.5-9B.
- For a single request, multi-token prediction (MTP) makes it about 1.6 times faster.
What it's for
- A fast sub-agent inside agentic systems: finding the root cause of a failing test, drafting and checking a fix.
- A general assistant for everyday questions and math and reasoning problems.
- Private deployment: run it locally or on your own infrastructure, keeping code and data under control.
Weights are on Hugging Face. JetBrains announced GGUF builds for llama.cpp, Ollama and LM Studio and the MTP head for vLLM as "coming soon"; a GGUF repository already appears on the company's Hugging Face page.
Why it matters
Small, fast models are becoming the "workforce" of agents: a large model plans, and several small models carry out tasks in parallel. With 2.5B active, Mellum2.1 fits on an ordinary GPU and cuts cost and latency, without sending code outside the company.
For businesses
- Test it as a sub-agent on well-scoped tasks (tests, small fixes), not as a replacement for frontier models.
- Compare on your own repository: vendor benchmarks don't always reflect your code.
- A good fit for sensitive code that can't leave the network.
Sources
- JetBrains blog (Oct 8, 2026): https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/
- Hugging Face, JetBrains/Mellum2.1-12B-A2.5B-Thinking: https://huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking
Want to run private coding agents on your own infrastructure? KGA IT offers a free consultation: kga-it.com.