Liquid AI Just Open-Sourced Its d1 Decision Models — Including an Experimental 600M Multimodal One
October 7, 2026
Liquid AI has released open-weight versions of its d1 decision models: d1-3B, which handles text and images, and d1-omni-600M, an experimental model that also takes audio. Why it matters: these are models that answer with a decision instead of a paragraph — and now developers can run them locally instead of calling an API.
What happened
Liquid AI, the MIT spinout led by CEO Ramin Hasani, posted two new checkpoints to Hugging Face on October 7: d1-3B and d1-omni-600M. The release follows the company’s October 5 announcement of an API-hosted d1 decision model, which had promised open weights for future versions. Two days later, they delivered.
What a “decision model” actually is
Most language models generate tokens even when the task is narrow — sorting a support ticket, or deciding whether a photo shows a defect. Decision models skip that. They process a state in a single forward pass and return a structured answer: a yes-or-no, a chosen label, or a rubric-based score. That makes them useful anywhere software needs an answer it can act on rather than a paragraph it has to parse — routing, inspection, classification, moderation, reranking.
The details of this release
d1-3B is built on Liquid AI’s LFM2.5-VL-3B vision-language model and accepts text plus images. d1-omni-600M combines the LFM2.5-Encoder-350M encoder with added vision and audio encoders, accepting text paired with images or audio. Liquid AI labels the smaller model experimental and says it remains under development. Both ship under the LFM 1.0 license, and the release includes sample code loading d1-3B through Transformers (with trust_remote_code=True).
On the numbers front: Liquid AI reports d1-3B answered a single question in 16 milliseconds on a Jetson AGX Thor, 26 ms on an AGX Orin, and 50 ms on an Orin Nano. It also reports an average of 74.1 across 11 public image benchmarks for d1-3B, just ahead of the 73.9 from its LFM2.5-VL-3B base model. Treat all of these as company-reported figures for now — no independent replication has been published, and the release gives no audio benchmark scores at all.
Why it matters
Here’s the subscriber’s take: this is a bet that a whole class of AI workloads doesn’t need a chatbot at all. A yes-or-no decision delivered in 16 milliseconds, running locally on an edge device, is a different product than an API call that returns three paragraphs. For routing and inspection tasks, that’s cheaper, faster, and more private. Liquid AI’s larger story since its founding has been models that fit inside real hardware limits, and d1 is that philosophy applied to structured decisions. The question worth watching: whether the experimental 600M audio-capable model matures, since audio decisions are the least proven part of the release.
FAQ
What is Liquid AI’s d1 model?
d1 is a decision model — instead of generating text, it returns structured outputs like yes-or-no decisions, label selections, or rubric-based scores, in a single forward pass.
What’s the difference between d1-3B and d1-omni-600M?
d1-3B handles text and images and is built on Liquid AI’s LFM2.5-VL-3B. d1-omni-600M is smaller, experimental, and also handles audio — but Liquid AI says it remains under development.
What license do the open weights use?
Both Hugging Face model cards list the LFM 1.0 license. Teams evaluating deployment should review those terms for their intended use.
How fast are these models?
Liquid AI reports d1-3B answered a single question in 16 milliseconds on an Nvidia Jetson AGX Thor. These are company-reported figures, not independent test results.
Sources: RuntimeWire; Liquid AI’s Hugging Face release announcement.