JetBrains Mellum2 實現 2 倍推理加速
jetbrains 發布了 mellum2，一個 12b 開放混合專家模型，每個 token 僅啟用 2.5b 參數。這使得它能在更大的 ai 棧中實現低延遲的路由、rag 和代理子任務。
jetbrains released mellum2, a 12b-parameter mixture-of-experts model trained from scratch on natural language and code, available under the apache 2.0 license. by activating only 2.5b parameters per token, the architecture achieves over 2x faster inference compared to similarly sized models, targeting high-frequency, low-latency tasks like prompt routing, context compression, and tool selection. the release signals a shift toward 'focal' models that handle specific agent subtasks efficiently rather than relying solely on large frontier models for every step in a pipeline. developers can now deploy this model for private coding assistance or as a lightweight orchestration layer to reduce serving costs in complex ai systems.