LFM2.5-2.6B評測:Liquid AI免雲端邊緣AI代理模型 | LFM2.5-2.6B Review: Liquid AI's On-Device Agent Model
By Kit 小克 | AI Tool Observer | 2026-08-08
🇹🇼 LFM2.5-2.6B評測:Liquid AI免雲端邊緣AI代理模型
LFM2.5-2.6B是Liquid AI在2026年8月4日推出的免雲端AI代理模型,只有26億參數、不到2.5GB就能塞進手機甚至樹莓派執行,不必連網、不必租GPU伺服器,就能規劃任務、呼叫工具、完成多步驟操作。對做邊緣運算、資安敏感產業,或單純想省API費用的開發者來說,這是最近最值得關注的邊緣AI代理模型之一。
什麼是LFM2.5-2.6B?
LFM2.5-2.6B是一款專為代理任務(agentic workload)設計的小型語言模型,用約34兆token訓練而成,具備128K上下文視窗與工具呼叫(tool calling)能力,權重開源,下載後可直接在本機硬體執行,不依賴雲端API。
效能實測:贏過4倍大的模型
Liquid AI公布的ToolSandbox基準測試中,LFM2.5-2.6B拿下77.83分,勝過參數量是它4倍的Qwen3.5-9B(76.44分)。在Apple M5 Max上跑出220 tokens/秒,Ryzen CPU上有113 tokens/秒,一般手機也能有30 tokens/秒——代表就算沒有獨立GPU,一般消費級裝置也能跑出堪用的AI代理模型速度。
LFM2.5-2.6B可以在哪些裝置上跑?
從筆電、手機到樹莓派這種入門級單板電腦,都能執行LFM2.5-2.6B,不需要雲端GPU算力。這對邊緣機器人、離線工具、或需要把敏感資料留在本機的金融、醫療產業特別實用。
為什麼免雲端AI代理模型對企業重要?
不用把資料送上雲端,就少一層外洩風險,也少一筆API帳單。對受監管產業(醫療、金融、政府)來說,LFM2.5-2.6B這類本機執行的邊緣AI代理模型,等於用一點效能換取資料主權與成本可控性。
小克實測心得
老實說,「小模型打敗大模型」的宣傳每隔幾個月就會出現一次,多半只贏在特定基準測試,換個任務就現形。LFM2.5-2.6B的亮點不在「打贏Qwen3.5-9B」,而在2.5GB塞進樹莓派還能跑代理任務這件事本身——如果你的需求是離線工具呼叫、本機自動化,這種等級的效能已經堪用;但如果要做複雜推理或長文生成,還是乖乖用雲端大模型比較實在。開源權重讓你可以自己下載測試,不必只看官方公布的基準數字。
🇺🇸 LFM2.5-2.6B Review: Liquid AI's On-Device Agent Model
LFM2.5-2.6B is the on-device AI agent model Liquid AI released on August 4, 2026 — a 2.6-billion-parameter model that fits in under 2.5GB and runs on a phone or even a Raspberry Pi, with no cloud connection or GPU rental required to plan tasks, call tools, and complete multi-step workflows. For developers working on edge computing, security-sensitive industries, or anyone trying to cut API bills, this is one of the more interesting on-device agent models to show up this year.
What Is LFM2.5-2.6B?
LFM2.5-2.6B is a small language model built specifically for agentic workloads, pre-trained on roughly 34 trillion tokens with a 128K context window and native tool-calling support. Its weights are open, so you can download it and run it locally without depending on a cloud API.
Benchmark Results: Beating a Model 4x Its Size
On Liquid AI's published ToolSandbox benchmark, LFM2.5-2.6B scored 77.83, edging out Qwen3.5-9B (76.44) despite having a quarter of the parameters. It hits 220 tokens/sec on an Apple M5 Max, 113 tokens/sec on a Ryzen CPU, and 30 tokens/sec on a typical smartphone — meaning consumer-grade hardware without a dedicated GPU can already run a usable agent model.
What Devices Can Run LFM2.5-2.6B?
Laptops, phones, and entry-level single-board computers like the Raspberry Pi can all run LFM2.5-2.6B without cloud GPU compute. That makes it practical for edge robotics, offline tools, or regulated industries like finance and healthcare that need to keep sensitive data on-device.
Why Does an On-Device AI Agent Model Matter for Businesses?
Keeping data off the cloud means one less leak vector and one less API bill. For regulated industries, a local-running on-device agent model like LFM2.5-2.6B trades a bit of raw performance for data sovereignty and predictable cost.
Kit's Honest Take
"Small model beats big model" headlines show up every few months, and most only win on one cherry-picked benchmark before falling apart on a different task. The real story with LFM2.5-2.6B isn't "it beat Qwen3.5-9B" — it's that a 2.5GB model running on a Raspberry Pi can handle agentic tasks at all. If your use case is offline tool-calling or local automation, this tier of performance is genuinely usable. For complex reasoning or long-form generation, you'll still want a full cloud model. Since the weights are open, you can download it and test it yourself instead of trusting the official benchmark numbers.
Sources / 資料來源
- VentureBeat: No cloud, no GPUs, no problem — Liquid AI's LFM2.5-2.6B
- MarkTechPost: Liquid AI Releases LFM2.5-2.6B On-Device Agentic Model
- Developers Digest: LFM2.5-2.6B Runs at 220 Tokens/s in Under 2.5 GB
常見問題 FAQ
LFM2.5-2.6B是什麼?
LFM2.5-2.6B是Liquid AI推出的26億參數開源AI代理模型,體積不到2.5GB,可在手機、樹莓派等裝置離線執行,具備128K上下文與工具呼叫能力。
LFM2.5-2.6B跟大型雲端模型比效能如何?
在ToolSandbox基準測試中,LFM2.5-2.6B以77.83分勝過參數量4倍的Qwen3.5-9B(76.44分),但複雜推理與長文生成仍建議使用雲端大模型。
LFM2.5-2.6B可以在哪些裝置上執行?
筆電、手機、Apple M5 Max、Ryzen CPU甚至樹莓派都能執行,不需要雲端GPU或網路連線。
免雲端AI代理模型對企業有什麼好處?
資料不用上傳雲端,降低外洩風險並省下API費用,特別適合金融、醫療等受監管產業。
延伸閱讀 / Related Articles
- Qwen3.8-Max評測:阿里2.4兆參數模型開源迎戰Fable 5 | Qwen3.8-Max Review: Alibaba's 2.4T Model Rivals Fable 5
- GPT-5.6 Luna降價80%評測:OpenAI引爆AI價格戰 | GPT-5.6 Luna Price Cut Review: OpenAI's AI Price War
- CrowdStrike威脅報告評測:AI攻擊暴增89%、29分鐘破網 | CrowdStrike 2026 Threat Report: AI Attacks Surge 89%
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言