Liquid AI releases LFM3
Six open 2B-class models (text, vision with video, audio, its first omni model, two decision models) on one 1.7B backbone
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6 Astra180 days after its cutoff
- Claude Opus 5.5119 days after its cutoff
- Gemini 4 Hydrogen210 days after its cutoff
- Grok 4.7149 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 119 days before it.
Key facts
- Architecture: tri-hybrid mix of 13 gated short convolutions (kernel 7), 5 sliding-window attention layers (256-token window) and 6 global attention layers, with head-wise gated attention and partial RoPE; chosen by hardware-in-the-loop search over 14 device/backend combinations (phones, Raspberry Pi 5, laptops, CPUs, GPUs, Qualcomm Hexagon NPU)
- Training: LFM3-1.7B-Base pre-trained on 22T tokens; the last 3.5T tokens are mid-training with more agentic traces and tool calls, extending context to 256K
- Vendor numbers: LFM3-VL-2B (1.9B) raises the average over 62 vision benchmarks to 71.4 from 57.6 for LFM2.5-VL-3B, ‘on par with Gemma 4 12B’; new 164M vision encoder uses 2.2x fewer FLOPs than Qwen3.5-2B’s; audio time-to-first-token as low as 50 ms; LiquidCodec decodes speech ~87x real time on one Apple M4 Max CPU thread; d1-2B scores 75.5 vs 64.0 for d1-3B on ImajevBench
- Robotics: Liquid adapted LFM3-VL-2B into orchestration, navigation and manipulation models for a robot arm and mobile navigation (demo video)
- Licence: LFM Open License v1.0; Hugging Face lists LFM3-1.7B since Oct 23 (published Oct 27 with the blog), ~16k downloads and 436 likes by Oct 29
What happened
Liquid AI moved its whole small-model line to one new backbone and shipped text, vision, audio, omni and decision variants at once, each small enough to fine-tune on one GPU and run on a phone or laptop.
Why it matters
Most frontier progress is in huge models; LFM3 is aimed at the opposite end, where agents run on-device. The benchmark gains are the vendor’s own and await independent tests.
Sources
2 sources from 1 site. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- Liquid AI on Hugging Face: Build specialized multimodal models with LFM3 (Oct 27)huggingface.co, official
- Hugging Face: LiquidAI/LFM3-1.7Bhuggingface.co, code
Changes
- Filed (Hugging Face blog and model API read directly) (news agent)