Liquid AI releases d1-3B and d1-omni-600M for local use

Liquid AI's open d1 models are positioned as freely downloadable, locally runnable versions of its decision-model line, which the company first introduced at the end of September 2026. Decision models work by selecting from a set of predefined choices and returning a confidence score, making them suited to tasks such as choosing an AI model for a goal or sorting many websites into categories. The company says these models run faster and cheaper than LLMs, which has drawn attention to the category since Jev appeared in September 2026. By releasing d1-3B and d1-omni-600M as open models, Liquid AI is making that approach available for local execution on consumer hardware rather than only through an API. The two models are based on the LFM2.5-VL-3B vision-language model and the LFM2.5-Encoder-350M encoder model respectively, with d1-omni-600M adding audio input. Performance charts in the company's post show they trail Liquid AI's d1 and Jev, but lead among open models of the same size. User reports of running d1-3B on a single GeForce RTX 3060 or RTX 3090 suggest the models are practical for local use, with one user describing speeds far faster than OpenAI's Decisions API. The availability of free, locally runnable decision models may appeal to developers who want to avoid API costs or keep processing on their own machines, though the body does not state any adoption or commercial results.