最新模型
Liquid AI Releases d1 Decision Model Weights for Text and Image Classification in a Single Forward Pass
On October 7, Liquid AI released d1-3B and the experimental d1-omni-600M, which return structured decisions without generating answers token by token. Local and Jetson deployment examples are available, but the low-latency figures depend on warm-up conditions, and evaluation results are still primarily based on the company’s own testing.
Liquid AI released weights for two d1 decision models on October 7. d1-3B accepts text and images; the experimental d1-omni-600M accepts text paired with images or audio. Both process the input state and question together, then read out answer probabilities in a single forward pass without producing output tokens. This gives applications that repeatedly classify, route, or make judgments an option for local deployment. Announcement
Architecturally, d1-3B is trained from the decoder-only LFM2.5-VL-3B; the 600M version uses a bidirectional text encoder followed by vision and audio encoders. Releasing the weights lets developers deploy and adapt the models themselves, moving decision steps onto edge devices. Architecture overview
The interface offers three answer types: noul returns a yes/no probability, choice selects a named option, and score rates items on an ordered scale. Multiple questions can share the same input; for example, a customer support ticket can be assessed at once for refund intent, responsible department, and urgency. The engineering benefit is avoiding text generation and answer parsing, though questions and candidate answers still need to be defined in advance. Jetson deployment guide
Speed figures should be read alongside their input conditions. The guide cites Liquid’s warm-up tests: on Jetson AGX Thor, one question took 16 milliseconds and three questions on the same input took 20 milliseconds; on Orin Nano, one question took 50 milliseconds. The model card also lists 8 milliseconds for one question on an RTX 4090, using BF16, the median of 20 measurements, and CUDA graphs; without that compilation path, the figure was 16 milliseconds. These numbers should not be treated as end-to-end camera pipeline latency. Deployment guide, performance conditions
For evaluation, d1-3B scored 48.57 on Decision Index 0.2.1, based on Liquid’s use of the official evaluator; this was not a formal leaderboard submission. For model selection, teams should still validate on Traditional Chinese data, check for class imbalance, and assess probability calibration. They should also confirm the terms of the LFM Open License v1.0 before deployment. Independent replication, along with latency and error rates on real-world input lengths, will be worth watching. Model card