JetBrains Releases Mellum2.1 A Fast Open-Source Coding Model

JetBrains has released Mellum2.1, an open-source model designed specifically for coding agents and fast sub-agents that run on your own hardware ¹. The model is a 12-billion-parameter Mixture of Experts thinking model, meaning only 2.5 billion parameters are active at any given moment . This makes it small enough to host privately while still being fast.

The release is available under the Apache 2.0 license on Hugging Face ¹. The architecture is unchanged from Mellum2, which JetBrains open-sourced in June 2026, but the upgrade comes almost entirely from reinforcement learning — a training method where the model improves by trying tasks and receiving rewards ².

Built for Agents That Actually Work

Mellum2 was fast, but it could not effectively work inside a real software repository ². After a summer of reinforcement learning in real environments, with millions of sandboxed runs across thousands of environments, Mellum2.1 can now explore a codebase, edit files, and check its own changes .

The biggest improvements are in agentic coding tasks. On SWE-bench Verified, the score jumped from 2.0 with Mellum2 to 47.0 with Mellum2.1 ¹. On SWE-bench Pro, it went from 0.0 to 28.0 . These benchmarks test how well a model can fix real bugs in real open-source projects.

JetBrains added reinforcement learning tasks in math, competitive programming, science, tool use, and software engineering ². The company filtered open datasets to remove broken tests, unverifiable answers, and tasks that were too easy or impossible for the model .

Speed and Performance

Mellum2.1 leads its peer group on several coding benchmarks ¹. It scored 82.0 on LiveCodeBench v6, compared with 75.4 for Qwen3.5-9B and 69.4 for Gemma 4 E4B . It also scored highest on HumanEval+ at 91.5 and BFCL v4 at 62.3 .

Qwen3.5-9B still outperforms Mellum2.1 on SWE-bench Pro at 38.0 versus 28.0, as well as on knowledge-heavy tasks like GPQA Diamond ¹. The numbers are self-reported by JetBrains and measured using the same evaluation pipeline for all models .

On speed, post-training did not change the architecture, so Mellum2.1 is as fast as its predecessor ². JetBrains reports it serves nearly twice the tokens of Qwen3.5-9B on a single NVIDIA H200 GPU under heavy load ¹. For individual requests, multi-token prediction makes it about 1.6 times faster .

Running It Yourself

The full model can be served on vLLM, and GGUF builds are available for local use through llama.cpp, Ollama, and LM Studio ¹. The smallest quantized version starts at 7.0 GB, making it practical to run on consumer hardware .

JetBrains recommends a temperature of 0.6 and a context length of 131,072 tokens ¹. The model ships in bfloat16 weights and supports tool calling for agents that need to interact with external systems .

Key facts

• Mellum2.1 is a 12B Mixture of Experts model with 2.5B active parameters
• SWE-bench Verified score jumped from 2.0 to 47.0 after reinforcement learning
• Released under Apache 2.0 license on Hugging Face
• Runs locally starting at 7.0 GB with GGUF quantization

Sources

1. JetBrains Releases Mellum2.1: A 12B MoE Open Model for Coding Agents — marktechpost.com
2. Mellum2.1 Gets to Work: A Fast Open Model for Coding Agents
3. Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains
4. Mellum2: JetBrains 12B MoE Code Model Guide - codersera.com — codersera.com
5. Mellum2 Goes Open Source: A Fast Model for AI Workflows