LiquidAI releases d1-3B — a decision model that answers in a single pass without generating text

On October 5, 2026, the model LiquidAI/d1-3B with the image-text-to-text task was published on HuggingFace. It is a decision model with 3.1 billion parameters, built on the LFM2.5-VL-3B base: the input is a state — text, JSON, images, or a mix of them — plus a set of questions, and the model returns typed answers in a single forward pass, without generating output tokens. The author claims 8 ms per decision on an NVIDIA RTX 4090, 9 ms on an AMD MI325X, and 30 ms on an Apple M5 Pro.

What it means

The largest weights file is model.safetensors at 5.8 GB, in safetensors format. It runs via transformers. The site already has 13 quantized builds in GGUF, MLX, and NVFP4 formats; the smallest working one is AtomicChat/d1-3B-GGUF in IQ4_XS quantization, taking up 1.5 GB. License — LFM Open License v1.0: you may use, modify, and distribute the model, including commercially, with annual revenue up to 10 million US dollars; for non-profit organizations the threshold does not apply. When distributing, you must include the license, note changes, and retain authorship notices. The model targets tasks that require choosing among options, a yes/no answer, or scoring: routing, triage, checks in pipelines. For local running on a CPU or a GPU with limited memory, the quantized GGUF builds are of practical interest — from 1.5 GB, which is noticeably below the original 5.8 GB. More on local AI tool news — in the CPU3D news feed.