Jetbrains · v2 · siet 2026-06-02 · 5× · tolest 29. Juni 2026
Mellum2 is an open-weight 12-billion-parameter language model from JetBrains built on a Mixture-of-Experts (MoE) architecture that activates only 2.5 billion parameters per token (64 experts, 8 active). It is the successor to the original Mellum (4B dense, code completion only) and extends coverage to code generation and editing, debugging, multi-step reasoning, tool use/function calling, and agentic coding workflows. Released under the Apache 2.0 license, model weights (Base, Instruct, and Thinking checkpoints) are freely available on Hugging Face. JetBrains explicitly positions Mellum2 as a "focal model" for high-frequency, latency-sensitive tasks within larger AI systems, not as a replacement for frontier models.
Features
| Base Model | Trained-from-scratch MoE model (12B total, 2.5B active, 64 experts/8 active per token) |
| Context Size | 131,072 tokens (≈128K); achieved through layer-selective YaRN scaling after pre-training. Architecture combines sliding-window attention (on 3 of 4 layers) with full-attention layers. |
| License | Apache 2.0 (open weights) |
| Platform | Hugging Face (open weights), local via vLLM/Transformers, cloud or self-hosting, JetBrains AI Assistant |
| Price | Free (open weights, Apache 2.0); users only pay for their own compute when self-hosting |
| Release Date | June 1, 2026 |
| Interface (IDE/CLI/Web) | IDE (JetBrains AI Assistant), CLI/API via vLLM (OpenAI-compatible), RAG/agent pipelines |
| Supported Languages | English and all major programming languages (general chat-interface usage) |