重點摘要
- Liquid AI 是一家源自 MIT CSAIL 的衍生公司,其研究基礎為液體時間常數網路(liquid time-constant networks),這是一種輸入依賴的循環架構,已獲 AAAI-21 接收。
- LFM2 模型混合了 10 個雙門控短卷積區塊與 6 個分組查詢注意力區塊,以減少記憶體使用並加快 CPU 推論速度。
- Liquid AI 報告稱,在自行選擇的硬體與執行環境下,其 CPU 上的預填充(prefill)與解碼(decode)速度比同規模模型快達 2 倍。
- LFM2 系列涵蓋 3.5 億至 83 億個參數,提供開放權重、32K token 的上下文窗口,以及視覺、音訊與檢索變體。
- LFM 授權對研究用途及營收低於 1,000 萬美元的企業免費,而規模較大的企業則需取得商業授權。
- AMD 於 2024 年 12 月領投了 Liquid AI 的 2.5 億美元 A 輪融資,根據 TechCrunch 報導,該公司估值超過 20 億美元。
深入探討
Liquid AI is an American AI company spun out of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT. TechCrunch names robotics researcher Daniela Rus as a co-founder, and the company grew out of liquid neural network research by Rus and her colleagues. The company releases and sells Liquid Foundation Models, or LFMs: general-purpose generative models built to be small and fast enough to run on edge devices, meaning hardware at the point of use such as smartphones, laptops, cars and robots, rather than on large server clusters. A foundation model is a model trained on broad data that can then be adapted to many tasks.
The company's name comes from its research roots. In a 2020 paper accepted to AAAI-21, Hasani, Lechner, Amini, Rus and Radu Grosu introduced liquid time-constant networks. These are recurrent neural networks, models that process a sequence one step at a time while carrying a hidden state, whose neurons follow simple differential equations. The twist is that each neuron's time constant, which controls how quickly it reacts to new input, is itself changed by a learned gate that depends on the input. The authors showed the networks stay stable and bounded and reported better time-series prediction than classical and modern recurrent networks. TechCrunch described the idea as inspired by the brain of a roundworm and noted that such networks can be much smaller than conventional models.
The commercial LFMs are not literal copies of those 2020 networks. They borrow the broader idea of input-dependent operators. Liquid AI's July 2025 LFM2 announcement frames its design space around what it calls linear input-varying (LIV) operators: linear operations whose weights are generated on the fly from the input they act on. Under that view, convolutions, recurrences and attention are all special cases, so an automated search can mix them. The company built a neural architecture search engine called STAR to explore that space, and for LFM2 it scored candidates by measured peak memory and prefill and decode speed on Qualcomm Snapdragon phone CPUs instead of relying on theoretical proxies such as cache size.
The search produced a hybrid backbone. Each LFM2 model in the first release has 16 blocks: 10 double-gated short convolution blocks and 6 grouped-query attention blocks, with a SwiGLU feed-forward layer and RMSNorm normalization in each block. A short convolution blends each token only with a few neighbors, so its cost grows linearly with text length and it needs no growing memory cache. Grouped-query attention, a cheaper variant of standard attention in which several query heads share one set of keys and values, handles the long-range lookups that convolutions cannot. Keeping attention to a minority of blocks is what shrinks memory use and speeds up generation on CPUs. The blog says the choice of short convolutions follows from the target hardware, embedded system-on-chip CPUs, and from how well existing kernel libraries run them.
The milestones are recent. Liquid AI released its first LFM series on 30 September 2024: LFM-1B (about 1.3 billion parameters), LFM-3B (about 3.1 billion) and LFM-40B, a mixture-of-experts model with 40.3 billion total parameters and roughly 12 billion active per token, all with a 32,000-token context window. That generation is now listed as deprecated. In December 2024 the company raised a $250 million Series A led by AMD, which TechCrunch reported at a valuation above $2 billion citing Bloomberg, with plans to tune models for AMD GPUs, CPUs and accelerators. LFM2 followed on 10 July 2025 with open-weight 350M, 700M and 1.2B checkpoints, and the LFM2 technical report posted to arXiv on 28 November 2025 extended the family to a 2.6B dense model and an 8.3B mixture-of-experts model with 1.5B active parameters. Newer LFM2.5 checkpoints, such as LFM2.5-1.2B-Instruct, keep the same 16-block layout but are pretrained on 28 trillion tokens and add multi-stage reinforcement learning.
Compared with the obvious alternatives, small transformer models such as Qwen3, Llama 3.2 and Gemma 3, the trade is speed and memory against breadth. Liquid AI's own table shows LFM2-1.2B scoring 55.23 on MMLU, a general-knowledge test, against 59.11 for the larger Qwen3-1.7B and 46.6 for Llama-3.2-1B-Instruct, while beating all three on IFEval, an instruction-following test, with 74.89. The company reports LFM2-700M decoding faster than Qwen3-0.6B in both ExecuTorch and llama.cpp despite being 16% larger. Compared with pure state-space models such as Mamba, which replace attention entirely with a recurrence, LFM2 keeps a few attention blocks for precise recall over long inputs.
The limitations matter. Nearly every benchmark and speed figure comes from Liquid AI itself, and its blog says it modified some evaluation procedures relative to the standard EleutherAI evaluation harness, so results are not directly comparable with other leaderboards. Qwen3 was tested with its reasoning mode turned off. Models of one to three billion parameters still hold far less world knowledge than frontier models and can be confidently wrong, so they suit narrow tasks such as extraction, classification and tool calling better than open-ended factual questions. Speed claims also depend on quantization: the published phone and laptop tests used 4-bit weights, which trade some accuracy for size.
In practice the requirements are modest. The models ship as open weights on Hugging Face with deployment packages for ExecuTorch, llama.cpp and vLLM, so a developer can run a 1.2B model on an ordinary laptop or phone without a GPU. The license is based on Apache 2.0 but adds a revenue threshold: academic and research use is free, commercial use is free for companies under $10 million in revenue, and larger companies must contact Liquid AI for a commercial license. Organizations that want custom models or enterprise support deal with the company directly.
03工作範例
在筆記型電腦上測量 LFM2 模型的速度
在沒有獨立 GPU 的筆記型電腦上安裝 llama.cpp(Liquid AI 支援的執行環境之一)。
從 Hugging Face 下載 LFM2-1.2B 的 4-bit GGUF 建置版本,以及類似大小的 Transformer 模型(如 Llama-3.2-1B-Instruct)。
在每個模型上使用相同的提示長度(例如輸入 512 個 token,輸出 128 個 token)以及相同的執行緒數,運行 llama.cpp 的基準測試工具。
記錄兩個模型的預填充速度(讀取提示時每秒處理的 token 數)和解碼速度(生成時每秒處理的 token 數)。
使用更長的提示(例如 4,096 個 token)重複測試,以觀察每個模型在上下文增長時的速度和記憶體使用變化。
向兩個模型詢問相同的十個實事和指令遵循問題,並標記哪些答案是正確的。
它顯示了什麼
一個包含兩個模型在同一台機器上的預填充速度、解碼速度和準確性的小表格。將其與 Liquid AI 發布的 2 倍 CPU 效能聲明進行比較,可以看出結果在多大程度上取決於晶片、執行環境和量化方式。
現實世界的實施
透過 llama.cpp 使用 Q4_0 量化在筆電 CPU 上完全離線執行 LFM2-1.2B,這是 Liquid AI 用於 AMD Ryzen HX370 上發布速度測試的兩個執行階段之一。
使用 ExecuTorch 在三星 Galaxy S24 Ultra 上進行基準測試,推出私密的手機端助理 Liquid AI,因此提示與回覆絕不會離開裝置。
使用 2025 年 11 月 LFM2 技術報告中所述的 LFM2-Audio,進行具備語音輸入與語音輸出獨立通道的即時語音對話。
在沒有專用 GPU 的硬體上,使用 LFM2-ColBERT(報告中的低延遲查詢與文件編碼器)建構多語言文件搜尋。
針對資料擷取等狹隘任務微調開源權重的 LFM2 檢查點,LFM2 授權允許年營收 1,000 萬美元以下的公司免費進行商業使用。
什麼是粉底液模型的未來?液態 AI LFM 指南
The direction Liquid AI has signalled is efficient models across more modalities rather than ever-larger dense networks. The November 2025 technical report already covers vision-language, audio and retrieval variants alongside an 8.3B mixture-of-experts model with 1.5B active parameters, and the LFM2.5 checkpoints push pretraining to 28 trillion tokens with reinforcement learning on top. Open questions remain. Independent evaluations of hybrid convolution and attention designs are still thin compared with those for transformers, so it is not settled how well the speed advantage holds across chips, runtimes and quantization levels beyond the ones the company tested. It is also unclear how far small models can close the knowledge gap with frontier systems, or whether their main role will be as fast local agents that hand harder requests to larger cloud models. The AMD partnership announced in December 2024 points toward hardware-specific tuning, which makes architecture search against real device measurements, the STAR approach, a likely area of continued work.
付諸實踐
為什麼這很重要
- 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
- 成本與預算. 商業條款和部署選項會影響長期成本和風險。
- 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。
留意
- 發佈公告可能會超過實際生產工作流程的穩定性。
- API 定價或政策轉變可能會在一夜之間打破假設。
- 單一供應商依賴性增加了鎖定和遷移成本。
後續步驟
- 使用您自己的任務和資料集評估提供者。
- 在整合之前查看隱私、安全和法律條款。
- 維護跨模型或供應商的後備計劃。
- 監控發行說明,以便路線圖的變更不會讓團隊感到意外。
資料來源與延伸閱讀
- arXiv (Liquid AI)LFM2 技術報告
- Liquid AI介紹 LFM2:市場上速度最快的裝置端基礎模型
- Liquid AI液態基礎模型:我們的第一個生成式 AI 模型系列
- arXiv (AAAI-21)液態時間常數網路
- TechCrunchLiquid AI 剛剛募資 2.5 億美元以開發更高效的 AI 模型類型
- Hugging FaceLiquidAI/LFM2.5-1.2B-Instruct
不斷探索
常見問題
什麼是粉底液模型?液態 AI LFM 指南?
Liquid Foundation 模型 (LFM) 是由 Liquid AI 建構的非 Transformer AI 模型,用於在手機、車輛和邊緣設備上實現高效的長上下文推理並降低記憶體使用量。
LFM2 技術報告中的「LFM」代表什麼意思?
該報告將 LFM2 擴展為液態基礎模型(Liquid Foundation Models)家族,並描述其模型變體。這個問題涉及報告的命名,而非生物學上的啟發。
根據 LFM2 技術報告,其混合骨幹網路結合了什麼架構?
2025 年 11 月的 LFM2 報告描述了一種緊湊的混合骨幹網路,使用閘控短卷積和少量的分組查詢注意力區塊;這是一種特定於模型的架構,而不是針對每個 LFM 的常數記憶體宣告。
LFM2-8.3B 混合專家模型的有效參數數量是多少?
技術報告列出了一個總參數 8.3B 的混合專家變體,具有 1.5B 的有效參數。這是該命名模型的配置事實。
LFM2 報告為該模型系列列出了什麼上下文長度?
LFM2 報告為該系列列出了 32K 的上下文長度。其他配置細節請檢查特定變體和部署套件。
LFM2 報告明確為哪種部署設定最佳化了其架構?
該報告將 LFM2 描述為專為裝置端延遲和尖峰記憶體限制下的裝置端部署而設計,並評估了智慧型手機和筆電等級的 CPU。這是一個設計目標和基準範圍,而不是在每個列出的裝置上部署的證明或隱私保證。