JetBrains releases Mellum 2.1 with open weights and a reported 47% SWE-bench score
The AMW Read
The reported benchmark gain at unchanged model size meaningfully updates JetBrains' coding-agent offering, while its open-weight strategy expands deployment options without proving production competitiveness.
JetBrains releases Mellum 2.1 with open weights and a reported 47% SWE-bench score
JetBrains released Mellum 2.1 on October 8 under Apache 2.0, according to the report. The coding model retains Mellum2's mixture-of-experts architecture: 12 billion total parameters and 2.5 billion active per token. JetBrains reports that SWE-bench Verified performance rose from 2.0% to 47.0%, attributing the improvement primarily to reinforcement learning in coding environments where the model explored repositories, executed commands, used tools, and checked edits. Thinking, Base, and Instruct variants are distributed through Hugging Face.
The market significance is an established developer-tools vendor offering a self-hostable model for coding agents alongside a market dominated by subscription APIs. Keeping the architecture unchanged makes this a concrete example of post-training improving repository-level task performance without increasing parameter count. That supports the case for specialized training as a competitive lever, although the reported benchmark gain alone does not establish production reliability or superiority to other coding systems. Open weights and commercial-use licensing also give builders more control over deployment and adaptation.
For engineering teams, the concrete next step is to evaluate Mellum 2.1 on their own repositories before making a deployment decision. The release includes GGUF builds for local inference through llama.cpp, Ollama, and LM Studio, creating a practical route to testing code workflows within company infrastructure. Buyers should compare task completion, tool-use reliability, latency, and total operating cost against their existing API setup; 2.5 billion active parameters alone does not establish workstation requirements or deployment economics.
