JetBrains has officially launched Mellum2.1, an open model specifically developed for creating intelligent coding agents and rapid sub-agents. This 12-billion-parameter mixture-of-experts (MoE) model, available under the Apache 2.0 license on Hugging Face, activates 2.5 billion parameters per token during operation.
While the underlying architecture remains consistent with its predecessor, Mellum2, the significant advancements in Mellum2.1 stem almost entirely from extensive reinforcement learning (RL) applied within real software environments. This rigorous training approach has resulted in a compact, self-hostable model capable of autonomously exploring code repositories, modifying files, and validating its own changes.
Architectural Insights and Enhanced Training
Mellum2.1, building upon the foundations of Mellum2 Thinking, functions as a reasoning model that articulates its thought process before generating an answer. Its architecture incorporates 28 layers and 64 experts, with a router mechanism selecting 8 active experts for each token. The model employs grouped-query attention, featuring 32 query heads and 4 KV heads. Notably, three out of every four layers utilize a 1,024-token sliding window, contributing to its substantial 131,072-token context length and a vocabulary size of 98,304 tokens. Model weights are supplied in bfloat16 format.
The development of Mellum2.1 heavily prioritized post-training efforts, shifting reinforcement learning from a brief final phase to a central component of the training process. JetBrains integrated a diverse array of RL tasks spanning mathematics, competitive programming, scientific applications, tool utilization, and software engineering. Rigorous filtering was applied to open RL datasets to eliminate faulty tests, unverified answers, and tasks deemed either too simplistic or unattainable. Crucially, for software engineering tasks, the model underwent training in live repositories, equipped with shell access and file-editing tools, receiving rewards for successfully passing tests. This involved launching millions of sandboxes across thousands of distinct environments.
Performance Benchmarks and Key Comparisons
Evaluations conducted by JetBrains, using a consistent pipeline in thinking mode, reveal Mellum2.1's performance against Mellum2, Qwen3.5-9B, and Gemma 4 E4B. A significant improvement was observed in agentic coding capabilities. The model's score on SWE-bench Verified dramatically increased from 2.0 to 47.0, while SWE-bench Pro saw a rise from 0.0 to 28.0. Terminal-Bench 2.1 performance also improved, climbing from 0.6 to 17.4. These agentic assessments utilized the open-source Pi v0.73.1 harness with a 114K-token context.
Mellum2.1 demonstrated leadership among the tested group in several benchmarks, achieving 82.0 on LiveCodeBench v6, 91.5 on HumanEval+, 79.4 on MBPP+, and 62.3 on BFCL v4. Despite these strengths, Qwen3.5-9B maintained its lead in categories such as SWE-bench Verified (50.0), SWE-bench Pro (38.0), AIME 25/26 (86.7), and GPQA Diamond (77.8). The model also exhibited improved safety, with its HarmBench score decreasing from 21.5 to 8.5, indicating better performance. It is important to note that benchmark scores can vary based on evaluation pipelines, with JetBrains' measurements for Qwen3.5-9B differing from vendor-reported figures.
Speed and Local Deployment Capabilities
Due to its unaltered architecture, Mellum2.1 maintains the same inference speed as Mellum2. Under heavy load on a single NVIDIA H200 GPU, JetBrains reports that Mellum2.1 processes nearly twice the number of tokens compared to Qwen3.5-9B. For individual requests, multi-token prediction (MTP) accelerates performance by approximately 1.6 times. Plans for an MTP head supporting vLLM speculative decoding are also in development.
Local deployment of Mellum2.1 is feasible, with options for serving the full model on vLLM, including specific configurations for reasoning and tool calling. JetBrains suggests particular inference parameters, such as a temperature of 0.6, top_p of 0.95, and top_k of 20. Furthermore, GGUF builds are underway, providing various quantization levels for CPU-based inference. These range from a reference BF16 build at 24.3 GB down to an MXFP4_MOE version at 7.0 GB, which is recommended for its balance of size and performance, offering an 85.6% top-token match.
Key Takeaways for Developers
- Mellum2.1 stands as a 12B MoE reasoning model with 2.5B active parameters, freely available under the Apache 2.0 license.
- Reinforcement learning within real code repositories significantly boosted its agentic coding prowess, evidenced by SWE-bench Verified climbing from 2.0 to 47.0.
- The model outperforms its tested counterparts on LiveCodeBench v6 (82.0) and BFCL v4 (62.3).
- While strong, it still trails Qwen3.5-9B in complex agentic software tasks and general knowledge benchmarks such as SWE-bench Pro, GPQA Diamond, and AIME.
- Compact GGUF builds, starting at 7.0 GB, render local execution of coding sub-agents a practical reality for developers.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
