edge-ai
3 篇文章
Gemma 4 登場:Google 說它用了和 Gemini 3 同樣的突破性技術
Google 發表 Gemma 4 開源模型家族,包含 31B Dense、26B MoE 和 E2B/E4B 邊緣模型,採 Apache 2.0 授權,支援 256K context、function calling、多模態,號稱在 Arena 上打贏 20 倍大的模型。
Paweł Huryn 稱:3B active parameters 的 Holo3 在 computer use 上贏過 GPT-5.4 和 Opus 4.6
Paweł Huryn 在 X 上稱,H Company 的 Holo3 在 computer use 任務上勝過 GPT-5.4 與 Opus 4.6,且僅有 3B active parameters。推文還稱它採 sparse MoE,並理論上可在單張 GPU 本地運行。
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。