手機就能跑推理模型?Liquid AI 把 LFM2.5-1.2B 壓進 900MB,邊緣 Agent 時代真的來了

Liquid AI 發布 LFM2.5-1.2B-Thinking:1.17B 參數、32K context,可在手機/NPU 裝置以不到 1GB 記憶體執行。官方數據顯示它在多數推理 benchmark 可匹敵或超越 Qwen3-1.7B,且速度更快、輸出 token 更少。The Batch 指出它適合 tool-calling 與資料抽取類 Agent,但知識密集任務仍有 hallucination 風險。