Mellum

Overview

Mellum is JetBrains’ family of fast language models, headlined by Mellum2 — a 12-billion-parameter open-source mixture-of-experts model built for real-time coding and language work. Where most teams reach for a giant frontier model and pay for every token, Mellum targets the opposite corner: sub-second latency, high throughput, and roughly half the inference cost of comparable models because only a fraction of its parameters activate per request. It is trained on more than 10 trillion tokens of permissively licensed code, web, and math data, then fine-tuned with SFT and reinforcement learning on verifiable rewards, so it behaves well on completion, next-edit suggestion, and short agent steps. The practical value is as a ‘workhorse’ model: smart routing between models, low-latency RAG pipelines, fast sub-agents inside a larger pipeline, and private local deployments where you cannot ship code to a third party. You can self-host the open weights or run it through JetBrains AI inside the IDE. It is not designed to be your only model — use it where speed and cost matter, and leave the heavy reasoning to a frontier LLM.

Key Features

• Open-source Mellum2 12B MoE model you can download and self-host
• Ultra-low-latency inference, often ~2x faster than similar-sized models
• Local or cloud deployment with full control over privacy and infrastructure
• Pre-trained on 10T+ permissively licensed tokens with RLVR post-training
• Purpose-built for code completion, next-edit, and fast agent sub-tasks

Pricing

模型开源免费自托管，托管版随 JetBrains AI 订阅；企业本地部署需商务授权，价格以官方为准。

Comparison

Against Claude Code, Cursor, and GitHub Copilot, the key distinction is that Mellum is a model, not an IDE or coding agent. Those three ship the editing experience; Mellum is the engine you can drop behind them for cheap, fast completions. If you want a turnkey assistant, the IDE tools win. If you are building your own agent or need on-prem code intelligence, Mellum is the piece to self-host.