JetBrains Releases Mellum2.1 Open Model for Coding Agents

JetBrains has released Mellum2.1, an open-source model specifically designed for coding agents and fast sub-agents. This 12 billion parameter mixture-of-experts (MoE) thinking model activates approximately 2.5 billion parameters per token. Mellum2.1 is available under the Apache 2.0 license on Hugging Face. The underlying architecture remains unchanged from its predecessor, Mellum2, with the significant upgrade stemming almost entirely from reinforcement learning (RL) conducted in real-world software development environments. This training approach has resulted in a compact, self-hostable model capable of exploring code repositories, editing files, and verifying its own modifications.

Mellum2.1 boasts a total of 12 billion parameters, with 2.5 billion active parameters per token, utilizing 64 experts where 8 are activated per token. It features an extensive 131,072-token context window. The model can be run on personal GPUs using frameworks like vLLM or SGLang, with speed tests conducted on an NVIDIA H200 GPU. For users preferring local execution, GGUF builds are available starting at 7.0 GB for use with llama.cpp, Ollama, and LM Studio. In terms of performance, Mellum2.1 surpasses Mellum2 on 15 out of 17 benchmark tests. It also demonstrates competitive performance against other models, winning 5 out of 17 benchmarks against Qwen3.5-9B. Specifically, Mellum2.1 achieved a score of 82.0 on LiveCodeBench v6, outperforming Qwen3.5-9B (75.4) and Gemma 4 E4B (69.4). However, on Terminal-Bench 2.1, it scored 17.4, trailing behind Qwen3.5-9B (21.7). The key advantage of Mellum2.1 lies in its strong coding capabilities and high throughput achieved with a relatively small number of active parameters, though it still lags behind Qwen3.5-9B in complex agentic software tasks and knowledge recall.

Mellum2.1 represents the latest iteration of Mellum2 Thinking, which JetBrains initially open-sourced in June 2026. The specific checkpoint released is Mellum2.1-12B-A2.5B-Thinking. This model is engineered as a reasoning model that explicitly outputs its chain of thought before providing an answer, making its decision-making process transparent. JetBrains has identified three primary use cases for Mellum2.1: as an agent worker, a general-purpose reasoning assistant, and for private, self-hosted deployments, catering to users who require enhanced data privacy and control. The model's architecture comprises 28 layers and 64 experts, with a router mechanism selecting 8 experts for each token. Its attention mechanism employs grouped-query attention, featuring 32 query heads and 4 KV heads. A sliding window attention mechanism, spanning 1,024 tokens, is implemented in three out of every four layers. The model supports a context length of 131,072 tokens and possesses a vocabulary size of 98,304 tokens. The model weights are distributed in bfloat16 format.

The training methodology for Mellum2.1 saw a significant shift in its post-training phase. The reinforcement learning component, previously a brief final stage, has now been integrated as the primary focus of the training process. JetBrains has incorporated additional data and techniques into this RL phase, aiming to enhance the model's ability to learn and adapt within simulated and real software development environments. This emphasis on RL in practical coding scenarios is intended to imbue Mellum2.1 with a deeper understanding of coding tasks and agentic behaviors, thereby improving its performance on benchmarks that simulate complex software development workflows. The company's commitment to open-sourcing these models, like Mellum2.1, aligns with a broader trend in the AI community towards fostering collaboration and accelerating innovation through accessible research and development tools.