FrontierThe story, in brief
JetBrains Releases Mellum2: A 12B MoE Model for Fast, Specialized Tasks in Multi-Model AI Pipelines
JetBrains just open-sourced Mellum2 — a 12B MoE model trained on 10.6T tokens. Here's why sparse models are becoming the new moat.
The KeyNews take
Why it matters
JetBrains enters the model release race with a specialized MoE (Mixture of Experts) architecture optimized for production AI pipelines. Open-sourcing under Apache 2.0 signals a shift toward modular, task-specific models competing against monolithic frontier models.
The key facts
5 to knowModel: Mellum2, 12B parameters, MoE architecture
Training data: 10.6 trillion tokens
License: Apache 2.0 (open source)
Use case: Fast inference, specialized tasks in multi-model pipelines
Company: JetBrains (traditionally dev tools, now model builder)
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: JetBrains releases Mellum2 under Apache 2.0 — a 12B MoE model trained on 10.6 trillion tokens for AI workflows.