Liquid AI ships LFM3: six 2B open-weight models, 1 GB each, built for phones and robots
AIFlow_ML · x · 2026-10-29
Liquid AI released LFM3 on Oct 27: six open-weight models of roughly 2B parameters, all branched from one base (LFM3-1.7B-Base), fine-tunable on a single GPU and runnable on phones, laptops, or robots.
- LFM3-1.7B: text-only, 256K context, tuned for local agents and tool calling
- LFM3-VL-2B: image/video input, pitched as "a strong foundation for robotics"
- LFM3-Audio-2B: voice; the family also covers decision models
The author, who deploys models on battery-powered hardware, argues this is the kind of release that matters for edge use. Self-reported numbers (not independently verified): LFM3-1.7B needs only 1.6 GB for a 128K KV cache vs Qwen3.5-4B's 4.3 GB; WebArena-Verified score of 38.42; LFM3-VL-2B averages 71.4 across 62 vision benchmarks, beating LFM2.5-VL-3B's 57.6. A robotics demo runs on a Jetson.
Related event: Liquid AI Releases LFM3 Family of 2B Open-Weight Models(5 posts)→
More from Infra
- US AI Power in Three Stages: 10GW+ Nuclear Contracts, Geothermal Live, SMRs Permitting — ingliguori · 2026-10-29
- Jarvislabs to Bring 1,024 NVIDIA B200 GPUs Online in December, Reservations Open — amaarora · 2026-10-29
- Lightpanda, a Zig-built browser for AI agents, uses 16× less memory than headless Chrome — techNmak · 2026-10-29
- 10T tokens/day of a rumored DeepSeek v4.1 flash on one GB300 NVL72? — _xjdr · 2026-10-29
- OpenBonsai open-sources a pipeline compressing BF16 models to ~2GB ternary GGUFs — ButtercupLyn100 · 2026-10-29
- KLA: Memory to hit 60% of process control revenue by H2 2027, CY27 WFE seen above $200B — zephyr_z9 · 2026-10-29