Liquid AI open-sources d1 decision models: 3B model answers in one forward pass, 8ms on RTX 4090
huggingface · x · 2026-10-08
Liquid AI released two open-weight "decision models" that produce an answer in a single forward pass instead of generating tokens:
- d1-3B: built on LFM2.5-VL-3B, takes text + image input, scores 48.57 on Decision Index v0.2.1 (public split), beating every model under 10B and matching the 12x-larger Decider 35B-A3B. It runs 8ms on an RTX 4090, 16ms on Jetson AGX Thor, and 50ms on the tiny Jetson Orin Nano — fast enough for real-time edge decisions.
- d1-omni-600M: an experimental checkpoint based on the LFM2.5-Encoder-350M bidirectional encoder with added vision/audio encoders, handling text+image or text+audio; scores 15.95 on the same index.
Training details: d1-3B uses weight averaging for a better base, then multiple fine-tunes with different seeds/data mixes merged again. The team says long-input training, shuffling answer options, and fixing data shortcuts mattered more than fancier techniques. Weights are on Hugging Face for local fine-tuning and deployment.
Related event: Liquid AI Open-Sources d1 Decision Models for Real-Time Edge Inference(9 posts)→
More from Infra
- Nebius up 160% vs CoreWeave's 10%: the AI infrastructure stock divergence explained — Beth_Kindig · 2026-10-08
- Microsoft's $5,999 Surface RTX Spark Dev Box preorders open, ships November — tomwarren · 2026-10-08
- Together scales open-source inference with IBM and NVIDIA on B300 cluster — togethercompute · 2026-10-08
- How a Self-Appending Summarizer Quadrupled Token Costs in a 1.4M-Conversation Agent — Good_Education4713 · 2026-10-08
- National Compute gifts $100M in compute credits to White House Genesis Mission — typewriters · 2026-10-08
- Ask: Is Qwen 3.8 Flash-Next Worth It Over 3.6 35B-A3B on 64GB RAM + 16GB VRAM? — msalsas · 2026-10-08