Liquid AI LFM3: Six 2B On-Device Models on One Backbone

Small language models keep getting more capable, and Liquid AI's new LFM3 family is a clear example. Released on October 27, 2026, LFM3 brings six on-device models covering text, vision, audio, combined "omni" input and decision-making onto a single shared backbone, all in the 2 billion parameter class. The pitch is simple: pick the model for your modality, fine-tune it on one GPU, and run it where your application lives.

• Six models, one base: LFM3-1.7B (text), LFM3-VL-2B (vision and video), LFM3-Audio-2B, LFM3-Omni-2B, plus the d1-2B and d1-omni-2B decision models
• Edge-sized: the text model fits in about 1GB at 4-bit precision and supports a 256K-token context window
• Efficiency claim: Liquid AI says LFM3-1.7B outperforms its LFM2.5-2.6B predecessor with 35% fewer parameters
• Single-GPU fine-tuning: every model in the family can be adapted on one GPU, according to Liquid AI

What Is New in the LFM3 Architecture?

LFM3-1.7B-Base uses what Liquid AI calls a tri-hybrid design across 24 layers: 13 gated short-convolution layers, 5 sliding-window attention layers and 6 global attention layers. Additions such as head-wise gated attention, which lets an attention head switch itself off, are meant to suppress the activation outliers that make quantization harder. The base model was pre-trained on 22 trillion tokens, with the final 3.5 trillion weighted toward agentic traces and tool calls.

Liquid AI designed the architecture with a hardware-in-the-loop search spanning 14 device and backend combinations, including phones, a Raspberry Pi 5, laptops, cloud CPUs, GPUs and a Qualcomm Hexagon NPU. That is why the company emphasizes long-context speed: it reports roughly 25% faster inference than LFM2-1.2B per parameter with 8K tokens in context.

How Good Are the LFM3 Small Models?

All benchmark figures below come from Liquid AI's own announcement, so treat them as vendor claims until independent testing arrives:

• Agents: LFM3-1.7B scores 38.42 on WebArena-Verified, versus 3.02 for LFM2.5-2.6B, and Liquid AI says it beats Qwen3.5-9B on long browser and business-app tasks
• Vision: LFM3-VL-2B lifts the average across 62 vision benchmarks from 57.6 to 71.4 compared with LFM2.5-VL-3B, which Liquid AI places on par with Gemma 4 12B
• Audio: LFM3-Audio-2B handles speech recognition, synthesis and translation across 18 languages, with time to first token as low as 50 milliseconds
• Omni: LFM3-Omni-2B, the company's first omni model, takes text, images, video and audio and can answer out loud

New encoders for vision and speech

Two in-house encoders do a lot of the heavy lifting. The 164M-parameter vision encoder uses about 3x less compute than the SigLIP2 encoder in earlier LFM vision models, and it halves video sequence length by merging pairs of frames. On the audio side, the new LiquidCodec tokenizer converts generated tokens into speech at about 87x real time on a single CPU thread of an Apple M4 Max.

Why On-Device AI Builders Should Care

The bigger idea is specialization. Instead of one large general model in the cloud, LFM3 encourages developers to fine-tune a compact model for one job, such as a voice assistant on a kiosk or a vision model on a robot, and keep the data local. Liquid AI says the text model is already post-trained for agent harnesses including Hermes Agent, Pi and OpenClaw, and all six models run through Hugging Face transformers.

If you followed our coverage of Liquid AI's LFM2.5-24B-A2B local agent model, LFM3 is the natural next step: smaller, more multimodal and easier to customize. Find more model releases in our AI news section.

Sources: Hugging Face blog: Build specialized multimodal models with LFM3 — October 27, 2026; Liquid AI news — October 27, 2026.