
JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP
JetBrains’ 12B mixture-of-experts model for agentic coding, tool use, and reasoning, with a 131K-token context window
Mellum2.1 Thinking is JetBrains’ open-weight reasoning model for coding agents and other complex tasks. It can work across a repository, use tools, run commands, and help diagnose and fix software issues. It also supports non-agentic coding, math, and reasoning tasks.
The model uses a mixture-of-experts architecture with 12 billion total parameters and 2.5 billion active parameters. Its MTP variant adds a speculative decoding head designed to improve generation speed. JetBrains reports a 1.4–1.8× decoding speedup with the head; performance varies by serving configuration and workload.
| Attribute | Value |
|---|---|
| Parameter Size | 12B-A2.5B |
| Context Length | 128k |
| Input Modalities | Text |
| Output Modalities | Text |
Use cases
Coding agents that investigate repositories, edit files, and check changes
Tool-using agents and specialized coding sub-agents
Difficult programming, math, and reasoning problems