Liquid AI's LFM3: 1.7B tri-hybrid backbone, 237 tok/s on M5 Max, open weights
RobertTLange · x · 2026-10-29
Liquid AI released LFM3, a 2B-class open-weight family (text, VL, audio, omni, decision models) on a new tri-hybrid 1.7B backbone.
- 24 layers: 13 gated short convs, 5 sliding-window (256) + 6 global GQA; only 6 layers hold KV (12 KiB/token, 256K context)
- 22T pretraining tokens + SFT, RL and 6-teacher on-policy distillation
- Wins 28/29 benchmarks vs Gemma 4 E4B, 14/29 vs the 5× larger Qwen3.5-9B
- 237 tok/s on M5 Max, 43.6 on iPhone 17 Pro (Q40); hardware-in-the-loop search covered 14 device/backend combos
- Positioned for 'specialized intelligence': fine-tunable locally into task experts
More from Infra
- Agents push compute back to CPUs: $221B server CPU market seen by 2030 — BenBajarin · 2026-10-29
- AI server revenue forecast to hit $920B in 2027, pushing market past $1T — Beth_Kindig · 2026-10-29
- Oracle deploys Google TPUs for the first time and scales Vera Rubin NVL72 clusters to multi-gigawatt scale with Nvidia — Beth_Kindig · 2026-10-29
- Google Cloud expected to post 95.1% Q3 growth, outpacing AWS and Azure — Beth_Kindig · 2026-10-29
- lium rents NVIDIA B300s at $4/GPU-hour, H200 prices for Blackwell Ultra — bittingthembits · 2026-10-29
- Microsoft Expected to Generate $8B+ Cash Flow in Q3 While Meta, Google, Amazon Burn AI Capex — Beth_Kindig · 2026-10-29