LFM2.5-24B-A2B評測:24B模型2B算力跑地端代理 | LFM2.5-24B-A2B Review: 24B Params, 2B Active, Runs Local
By Kit 小克 | AI Tool Observer | 2026-10-16
🇹🇼 LFM2.5-24B-A2B評測:24B模型2B算力跑地端代理
LFM2.5-24B-A2B 是 Liquid AI 在 2026 年 10 月 13 日推出的開源權重模型,總參數 24B、但每次推論只動用 2.45B 算力,目標是讓一般筆電也能跑 AI agent,不用把任務全部丟給雲端。這篇評測從架構、實測速度到真實落差,帶你看這顆「地端模型」到底好不好用。
LFM2.5-24B-A2B是什麼?
它是 Liquid AI 的 LFM2.5 系列目前最大、也是最後一顆模型,專門做成推理(reasoning-only)架構,回答前會先產生一段思考鏈。模型支援 131,072 token 的上下文窗口、16 種語言,詞彙表從上一代的 6.5 萬擴充到 12.8 萬,對非拉丁語系(包含中文)的斷詞效果更好。
為什麼2B算力能跑24B模型?
關鍵在 MoE(Mixture of Experts)架構:每個 MoE 區塊有 64 個專家,但每次只路由 4 個專家參與運算,所以硬碟和記憶體要吃下 24B 參數的空間,實際運算量卻只有 2.45B。這讓 LFM2.5-24B-A2B 在筆電等級硬體上也能維持不錯的生成速度,不需要資料中心等級的 GPU。
效能實測:筆電跑起來多快?
官方公布的數字相當具體:
- Apple M5 Max:每秒 184 token
- AMD Ryzen AI Max+ 395:每秒 114 token
- 單張 NVIDIA B300:每秒超過 21,000 output token
- 發布當天即支援 llama.cpp、MLX、vLLM、SGLang、ONNX,不用等社群補丁
LFM2.5-24B-A2B好用嗎?
工具呼叫(tool-use)和代理任務是它的強項:在 8 項工具呼叫與代理相關的基準測試中贏了 5 項,像 AutomationBench 拿到 27.69 分,是第二名的三倍以上,整體表現跟參數更大的 Qwen3.6-35B-A3B 打平。但誠實說,知識類題目跟程式能力明顯落後,如果你要的是寫程式或查冷知識的助手,它不是首選;它的強項很明確是「本地跑自動化任務」,不是全能型模型。
怎麼安裝跑LFM2.5-24B-A2B?
Liquid AI 把 base 和 post-trained checkpoint 都放上 Hugging Face,直接用 llama.cpp 或 MLX 載入即可在 Mac/Windows 筆電跑起來,不需要額外轉檔或等待量化版本,這點比很多開源模型友善不少。
好不好用,試了才知道。
🇺🇸 LFM2.5-24B-A2B Review: 24B Params, 2B Active, Runs Local
LFM2.5-24B-A2B is Liquid AI's open-weight model released on October 13, 2026 — 24B total parameters, but only 2.45B active per inference, built so a regular laptop can run AI agents locally instead of shipping every task to the cloud. This review walks through the architecture, real-world speed, and where it actually falls short.
What is LFM2.5-24B-A2B?
It's the largest and final model in Liquid AI's LFM2.5 family, built as a reasoning-only model that produces an explicit chain of thought before answering. It supports a 131,072-token context window, 16 languages, and a vocabulary expanded from 65,000 to 128,000 tokens for better non-Latin tokenization, including Chinese.
How does 2B active compute run a 24B model?
The trick is its Mixture of Experts (MoE) architecture: each MoE block has 64 experts, but only 4 get routed per token. So while disk and memory have to hold all 24B parameters, actual compute per forward pass is just 2.45B. That's what lets LFM2.5-24B-A2B hold reasonable generation speed on laptop-class hardware instead of requiring a data center GPU.
How fast does it actually run?
Liquid AI's published numbers are specific:
- Apple M5 Max: 184 tokens/sec
- AMD Ryzen AI Max+ 395: 114 tokens/sec
- Single NVIDIA B300: over 21,000 output tokens/sec
- Day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX — no waiting on community patches
Is LFM2.5-24B-A2B actually good?
Tool-use and agentic tasks are where it shines: it leads or matches on 5 of 8 tool-use/agentic benchmarks, scoring 27.69 on AutomationBench — more than triple the next-best score — and roughly matching the larger Qwen3.6-35B-A3B overall. Honestly though, it clearly trails on knowledge and coding benchmarks, so if you want a coding assistant or a trivia machine, this isn't it. Its lane is narrow but clear: local, on-device automation, not a do-everything model.
How do you run LFM2.5-24B-A2B locally?
Liquid AI published both base and post-trained checkpoints on Hugging Face, loadable directly via llama.cpp or MLX on Mac or Windows laptops with no conversion step or waiting for quantized builds — friendlier than most open-weight releases on day one.
You won't know if it's good until you try it. 好不好用,試了才知道。
Sources / 資料來源
- Liquid AI Releases Open-Weight LFM2.5-24B-A2B for On-Device Agents (Unite.ai)
- Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents (MarkTechPost)
- LFM2.5-24B-A2B: Liquid AI's Biggest Model, 131K Context
常見問題 FAQ
LFM2.5-24B-A2B是什麼模型?
Liquid AI在2026年10月發布的開源權重推理模型,24B總參數、2.45B活躍參數,採用MoE架構,專為本地跑AI代理設計。
跑LFM2.5-24B-A2B需要什麼硬體?
不需要資料中心GPU,一般筆電(如Apple M5 Max或AMD Ryzen AI Max+ 395)就能跑,llama.cpp、MLX當天即支援。
LFM2.5-24B-A2B適合用來寫程式嗎?
不太適合。它的工具呼叫與代理任務表現強,但知識與程式能力明顯落後,不是全能型模型。
LFM2.5-24B-A2B和Qwen3.6-35B-A3B比起來如何?
在工具呼叫與指令遵循基準上大致打平,但LFM2.5-24B-A2B活躍參數更少(2B對3B),效率更高。
延伸閱讀 / Related Articles
- RTX Spark評測:74B AI模型裝進筆電,官方沒附跑分 | RTX Spark Review: 74B AI Model, No Benchmarks Yet
- Claude假線報評測:AI誤報兇殺案,Anthropic緊急斷網 | Claude Fake Tip Review: AI Sent False Murder Tip to Police
- XBOW評測:GPT-5.6挖出Chrome零日,奪25萬美元賞金 | XBOW Review: GPT-5.6 Finds Chrome Zero-Day, Wins $250K
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