88Z AI, 9 October — JetBrains has released the open-source model Mellum2.1, featuring a 12B mixture-of-experts architecture. The model makes reinforcement learning central to training, supplements data across multiple domains, and launches millions of sandboxes. It offers capabilities for codebase exploration, file editing, and fault diagnosis, with inference throughput nearly double that of Qwen3.5-9B under high load and a speed increase of approximately 1.6x via multi-token prediction.
At a glance
- Uses a 12B mixture-of-experts architecture with reinforcement learning as the training core
- High-load throughput is nearly double Qwen3.5-9B, with a 1.6x speed increase
The update
JetBrains has officially launched Mellum2.1, an AI programming model built on a 12B mixture-of-experts architecture. Regarding training strategy, Mellum2.1 expands reinforcement learning from a final step to the core of the training process, supplementing it with professional data in mathematics, algorithm competitions, and software engineering. Millions of sandbox environments were initiated during training to support complex tasks. In terms of functionality, the model can deeply explore codebases, perform file editing operations, diagnose the root causes of test failures, and verify fixes. In terms of performance, its inference throughput under high load is nearly double that of the Qwen3.5-9B model. Additionally, by introducing multi-token prediction technology, the model's generation speed has increased by approximately 1.6 times, significantly optimizing coding efficiency.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- IT之家Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.