JetBrains releases Mellum2.1, a 12B open-weight model for coding agents

JetBrains has released Mellum2.1, an open-weight reasoning model designed for coding agents and fast subordinate workers. The company describes a 12-billion-parameter mixture-of-experts architecture with 2.5 billion active parameters per token. Its official model card lists a 131,072-token context window, bfloat16 precision and an Apache-2.0 license. A downloadable Hugging Face checkpoint makes artifact-level verification possible, rather than leaving developers with only a hosted API.

JetBrains says the base architecture is unchanged from Mellum2, opened in June, while the new work concentrates on post-training. Reinforcement learning covers mathematics, competitive programming, science, tool use and software engineering. The publisher describes millions of sandboxed experiments in real repositories, where the model can inspect code, edit files and receive feedback from executable tests. Those statements describe the publisher's training process; they are not an independent audit of the data or training runs.

The official model card reports 47.0% on SWE-bench Verified and 82.0% on LiveCodeBench v6. JetBrains also claims that, on its H200 serving setup, the model serves nearly twice as many output tokens as Qwen3.5-9B at high load. It attributes additional single-request acceleration to multi-token prediction. These results depend on the company's hardware, evaluation harness and settings. Open Model Weights has not independently reproduced the quality or throughput numbers.

The model card documents vLLM serving and reasoning-parser settings. JetBrains has separately published a GGUF repository with several quantizations, while additional speculative-decoding support remains on its roadmap. For the evidence registry, the next step is to pin exact repository revisions, inspect model-card claims against config and license files, and record BF16 and GGUF artifacts separately. Until then, this is a verified release announcement and a candidate for registry inclusion, not a newly field-verified OMW record.