Liquid AI has released two open-weight decision models, including a 3-billion-parameter vision-language system that the company says leads models under 10 billion parameters on its public benchmark.
Liquid AI has moved into the fast-growing market for decision-focused models with the release of two open-weight systems, one designed for text-and-image inputs and the other built for text, image and speech combinations. In a post published on 7 October, the company introduced d1-3B, a 3.12 billion-parameter model based on its LFM2.5-VL-3B vision-language architecture, and d1-omni-600M, a 587 million-parameter experimental multimodal checkpoint built on the LFM2.5-Encoder-350M backbone.
The launch comes as vendors race to capitalise on interest in decision models, a class of systems that has drawn attention since OpenAI’s JEV model appeared in September. Liquid AI is positioning itself as a new entrant in that field, and it has made both models available on Hugging Face, where users can download them, fine-tune them and deploy them locally.
According to the company, d1-3B scored 48.57 on the Decision Index v0.2.1 public test set, which Liquid AI says puts it ahead of every model under 10 billion parameters. The model is intended for text and image reasoning, while d1-omni-600M extends the company’s approach to combinations of text and image, or text and speech, in what Liquid AI describes as its first experimental multimodal decision-model checkpoint.
Liquid AI also highlighted latency figures to underline the models’ suitability for edge deployment. On an NVIDIA RTX 4090, the company says d1-3B answers a single question in 8 milliseconds and processes a 384-pixel image in 102 milliseconds. On Jetson AGX Thor, it reports 16 milliseconds per question, while on Jetson Orin Nano the figure rises to 50 milliseconds. Independent user tests cited by IT Home suggest the model can run on a GeForce RTX 3060 with 12GB of memory, taking about 25 milliseconds for a single judgement and around 146 milliseconds for a 640×480 image, with peak memory use of roughly 6GB. One user also reported 8-millisecond inference on an RTX 3090.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
