LiquidAI releases d1-2B — a decision model for text, images, and video

The model LiquidAI/d1-2B with the image-text-to-text task was published on HuggingFace on October 24, 2026. It is a decision model built on LFM3-VL-2B: it does not generate tokens but instead returns calibrated answers with probabilities in a single forward pass — it extracts text, numbers and dates, finds objects in images and estimates quantities as an interval. The author claims 5.7 ms per decision on NVIDIA H100 and 3.8 ms on AMD MI325X.

What it means

The model contains 1.9 billion parameters, and the largest weights file is 3.6 GB in safetensors format. It runs via transformers. The site already has five quantized builds in GGUF and MLX, the smallest working one being Q4_K_M at 1 GB. License — LFM Open License v1.0: the model may be used, modified and distributed, including commercially, as long as annual revenue does not exceed 10 million US dollars. For non-profit organizations the threshold does not apply. When distributing, you must include the license, note changes and retain attribution notices. For local running on your own hardware, the 1 GB quantized build looks feasible even for consumer GPUs, while the full 3.6 GB weights will fit into the VRAM of mid-range graphics cards. Context — 128K tokens, vocabulary — 128,000 tokens, languages include Russian, English, Arabic, Chinese, French, German, Italian, Japanese, Korean and Portuguese. The model is aimed at classification and structured-answer extraction tasks rather than text generation, so comparing it directly with regular text-to-video or chat models is not worthwhile. You can follow new tools for local running in the CPU3D news feed.