采用LFM3：专门情报的每一种模式

中文版正在随流水线自动整理，稍后刷新即可；可先阅读下方英文原文。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Introducing LFM3: Every Modality for Specialized Intelligence

Today, we’re excited to announce LFM3, our new generation of Liquid Foundation Models (LFMs) built for specialized intelligence. The LFM3 model family is multimodal by nature and brings text, vision, audio, omni, and decision models onto one new shared backbone. The first release of the LFM3 family covers models in the 2B parameter range.

Built on a new hybrid architecture and 22T tokens of pre-training, LFM3 is far more expressive than LFM2 while maintaining its efficiency. The extra capacity in a small, fast model makes LFM3 the ideal starting point in its class for specialized intelligence: this is because it can be fine-tuned locally, it can run on any device, and it is capable enough to become a reliable expert on a specialized task.

The base and post-trained models are open-weight and available today on Hugging Face, our Playground, and the Liquid AI API. Check out our docs on how to run and fine-tune them locally.

Download weightsTry in Liquid APITry in PlaygroundBenchmark in Pipette

New architecture for more expressivity

LFM2 showed that a hybrid of short convolutions and attention could run fast on-device without sacrificing quality. Its design came from our automated hardware-in-the-loop architecture search. For LFM3, we used this search to redesign that architecture for agentic use cases, which need more capability per parameter and fast inference over long contexts.

We also expanded the hardware in the loop, from CPUs in one phone and one laptop for LFM2 to 14 device and backend combinations for LFM3: phones (Samsung Galaxy S26 Ultra and OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max and AMD Ryzen AI), and Arm and x86 cloud CPUs, plus GPUs and the Qualcomm Hexagon NPU.

Gated short convolutions still make up over half the layers. To add capability without slowing down long contexts, we tried swapping some of them for other sequence blocks, and sliding-window attention (SWA) gave the best quality–speed tradeoff. We kept the other changes described below only where they improved quality without a meaningful speed cost. The payoff shows at long context. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices, and with 8K tokens in context it is about 25% faster.

Tri-hybrid layer mix. LFM3 combines three operators: gated short convolutions, sliding-window attention (SWA), and global attention. Of its 24 layers, 13 are convolutions, 5 are SWA with a 256-token window, and 6 are global attention. The convolutions also use a kernel size of 7, up from 3 in LFM2.

Head-wise gated attention. A sigmoid gate scales each head’s output, so a head can switch off when it has nothing to add. This can help suppress activation outliers that hurt quantization and improve training stability.

Partial RoPE. Global attention layers rotate only the highest-frequency half of each head and leave the low-frequency half unrotated, so those dimensions match on content regardless of distance. SWA layers keep full RoPE.

Sandwich norm. Each block normalizes both its input and its output, which helps keep activations well-scaled.

LFM3 uses the same 128K vocabulary as LFM2.5-8B-A1B, built to tokenize non-Latin scripts more efficiently.

Training. LFM3-1.7B-Base was pre-trained on 22T tokens. Compared with LFM2, the mix adds more scientific papers, repository-level code, a wider range of knowledge sources, and considerably more math. The final 3.5T tokens are a mid-training stage on curated, higher-quality data, with a larger share of agentic traces and diverse tool calls to prepare the model for agentic post-training. This stage also extends the context window to 256K tokens.

A multimodal family for every need

The first LFM3 release includes six models, all branched from the same LFM3-1.7B-Base backbone and all around 2B parameters. Four of them cover different input modalities (text, vision, audio, and omni), and the last two are decision models specialized from the vision and omni models. Instead of running one large general-purpose model, you can pick the one that fits your use case, customize it, and run it on-device.

For a deeper dive into the performance and inference results behind these models, refer to the respective model cards.

Text

LFM3-1.7B is a text model for local agents and specialized applications with a 256K-token context window.

It is post-trained to run in agentic harnesses such as Hermes Agent, Pi, and OpenClaw. It can triage your inbox, manage your calendar, research topics on the web, send you morning briefings on WhatsApp, and use tool search to find the right tool among thousands of options. Since tokens are free, dozens of agents can run continuously in the background. You can adapt it to your domains, harnesses, and tools with a short fine-tuning run on a single GPU.

On agentic tasks, LFM3-1.7B competes with models several times its size. It leads on instruction following and function calling on BFCL and ToolSandbox, and beats Qwen3.5-9B on long tasks in browsers and business apps (WebArena-Verified, AutomationBench). Larger models still win on agentic coding and knowledge, where size matters most, but search helps close the knowledge gap, nearly doubling its score on Humanity's Last Exam.

To achieve these results, the model went through four stages of post-training:

- Supervised fine-tuning (SFT): 1.75T tokens across various tasks, including chat, reasoning, tool calling, and agentic traces.

- Mixed reinforcement learning: many domains trained at once, from math and code to SQL and structured outputs. A length penalty in the reward keeps answers concise.

- Multi-domain on-policy distillation (MOPD): six specialist RL teachers (code, math, tool use, long context, agentic work, and enterprise workflows) are distilled back into a single model.

- Agentic RL: tasks from web browsing to office work are run inside real agent harnesses, preparing the model for deployment.

On-device, we measure metrics with llama.cpp at Q4_0 and a 32K-token context. LFM3-1.7B is the fastest model on all three devices, decoding 1.8 to 2.8 times faster than the next one, and it uses the least memory. Long contexts stay cheap: a full 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B.

On a single NVIDIA B300 with SGLang in BF16, LFM3-1.7B can produce 3.6 billion tokens per day at 1,024 concurrent requests, more than any other model in this comparison.

To show what this throughput makes possible, we ran more than 300,000 LFM3-1.7B agents in parallel to investigate the Enron email archive.

LFM3-1.7B can be served at low latency under random traffic, enabling real-time use cases like online fraud detection when run on Liquid RIPL. P95 request latency remains stable even under heavy load, outperforming competing models served on vLLM.

Vision

LFM3-VL-2B can watch, reason, and understand the physical world:

- Video understanding: for the first time, an LFM vision model takes video as input. Ask questions about a recorded lecture, a walkthrough filmed on your phone, or a clip from a camera feed.

- Thinking: the model reasons step by step before answering. This helps it better understand user inputs and solve harder visual questions such as math, charts, and multi-step problems.

- Spatial and embodied reasoning: the model’s ability to localize objects and reason about their spatial relationships provides a strong foundation for downstream robotics applications.

LFM3-VL-2B is a leap in performance over the LFM2.5-VL series. Across 62 vision benchmarks, LFM3-VL-2B scores significantly ahead of all models of similar size by at least 7.4 points, and despite its size, it matches the performance of Gemma 4 12B, a model more than 6 times larger, surpasses Gemini 3.5 Flash Lite and is within 2 points of GPT-5.6 Luna. For more details, see the model card.

Notes: Each category score is the unweighted mean of its benchmark scores. The Average row is the unweighted mean across all 62 evaluations.

All of this can be run efficiently on-device, which enables everyday use of intelligence without the data leaving your device or the need for a cloud connection. On-device, LFM3-VL-2B has the highest prefill throughput and lowest memory footprint across all measured devices, compared to other models of its size class.

On a single NVIDIA B300 with SGLang in BF16, LFM3-VL-2B can produce over 12,000 tokens per second, higher than any other model of the same size class.

LFM3-VL-2B can be served at low latency under random traffic, enabling real-time use cases like image asset tagging when run on Liquid RIPL. P95 request latency remains stable even under heavy load, outperforming competing models served on vLLM.

Vision encoder

In addition to improvements to the LFM3 backbone, LFM3-VL comes with our new in-house vision encoder which uses 3× less compute than the SigLIP2 encoder in previous LFM VL models.

The Liquid vision encoder combines convolution and attention to make it significantly more efficient than its ViT counterpart at the same quality. The performance-efficiency graph below illustrates this. Our 164M-parameter encoder matches Qwen3.5-2B’s vision encoder at 55.9% average accuracy across 13 vision-language model benchmarks, while requiring 2.2× fewer FLOPs.

For video, the encoder jointly embeds each pair of sampled frames into one temporal slice of visual features, halving the temporal sequence length before further encoding. Together, these choices make the encoder a foundation for both richer image understanding and efficient video processing.

Audio

LFM3-Audio-2B is a compact, fully featured audio model that runs locally on laptops and edge devices. It takes text or audio as input and responds with text, speech, or both.

LFM3-Audio-2B uses a new architecture with an audio encoder and codec developed in-house. It outperforms LFM2.5-Audio in speech quality and now supports 18 languages. New capabilities include audio understanding, general-purpose audio reasoning, and tool calling.

- Voice chat: ask questions aloud and receive streaming replies in text and speech. For text-only responses, the model can also reason before answering.

- Multilingual speech: transcribe, synthesize, and translate speech across 18 languages.

- Audio understanding and reasoning: ask the model to interpret recordings and answer questions about the speech, sounds, and events they contain.

- Tool calling: turn spoken requests into tool calls to control apps and devices. The model can also respond in text or speech alongside those calls.

ASR (WER, %, lower is better)

English TTS (UTMOSv2, 1-5, higher is better) (WER, %, lower is better)

VoiceBench, Tool calling, Audio understanding (audio in, text out, %, higher is better)

Our new LFM-based audio encoder is built from a mixture of sliding-window attention and short-convolution layers, keeping memory utilization bounded. The new encoder is fully causal, allowing the model to process live microphone input and perform prefill while you speak. This reduces the wait after you finish, with time to the first token as low as 50 ms.

LiquidCodec, our new audio tokenizer, powers multilingual speech generation in both LFM3-Audio and LFM3-Omni. Built on the LFM3 architecture, LiquidCodec turns generated tokens into speech at about 87× real time on a single CPU thread (Apple M4 Max), supporting streaming decoding.

Quantization compresses each 80 ms of speech into 12 discrete tokens and, in contrast to previous LFM Audio models, allows multiple tokens to be predicted iteratively in parallel rather than requiring sequential prediction. As the backbone generates new tokens, the detokenizer progressively converts them into speech, all faster than real-time so that conversations are smooth and responsive.

Omni

LFM3-Omni-2B is our first omni model, bringing vision, audio, and language together in a single model that can see, listen, and talk back:

- Spoken visual question answ

Auto-fetched and translated by AI Atlas. All rights belong to the original author.