JetBrains Releases Programming AI Model Mellum2.1, Delivering Twice the Throughput of Competitors Under High Load

JetBrains Releases Programming AI Model Mellum2.1, Delivering Twice the Throughput of Competitors Under High Load

AIbase

Well-known development tools vendor JetBrains has officially released its brand-new Mellum2.1 programming AI model, with deep enhancements focused on agentic coding capabilities. The model continues to use a 12B mixture-of-experts architecture and an open-source license, aiming to further lower the barrier to private deployment for enterprises and individual developers.

Reinforcement Learning Powers Agentic Capabilities

In terms of its training mechanism, Mellum2.1 expands reinforcement learning from what was previously a short final-stage step into the core of the training process, and has completed specialized reinforcement for software engineering and algorithms across millions of sandbox environments. The upgraded model has strong autonomous retrieval and verification capabilities, able to accurately pinpoint the root cause of failing tests and draft fixes.

Inference Performance Significantly Improved

Thanks to the introduction of multi-token prediction technology, the model's response speed in single-request scenarios improved by about 1.6 times. Official test data shows that, when facing high-load inference requests, the model's token throughput is nearly twice that of the popular open-source model of the same class, Qwen3.5-9B, offering excellent cost-effectiveness for large-scale code deployment.

This article comes from the AIbase Daily

Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day we bring you the hottest content in the AI field, with a focus on developers, helping you stay on top of technology trends and learn about innovative AI products and applications.