Liquid AI d1: Open Decision Models That Answer in 8 Milliseconds

Liquid AI d1 is the newest entry in the "decision model" trend, and this one is open weights. Liquid AI released d1-3B and d1-omni-600M on Hugging Face this week: models that do not write text at all, but return a typed answer in one pass, fast enough to run on a Jetson board at the edge.

What actually happened

Liquid AI, the company behind the LFM model family, published "Open d1" with two checkpoints. d1-3B reads text and images. d1-omni-600M is an early research release that reads text with an image or text with audio. Both are on Hugging Face, and Liquid says they have day-one llama.cpp support.

The key twist: they produce zero output tokens. MarkTechPost says every response reports "output_tokens: 0." Instead of generating a sentence, the model returns a calibrated, structured answer in a single forward pass. Think "yes or no," "which team owns this ticket," or "how urgent is this," not chat.

Liquid lists use cases like routing, content moderation, intent classification, reranking, agent guardrails and visual inspection. It also built ten live-camera demos, from gesture games to moderation, that you can try in its Hugging Face space.

The numbers, with context

On Liquid's seven public text benchmarks, d1-3B averages 82.9, ahead of Decider 4B at 81.1. The smaller d1-omni-600M averages 78.4, beating Decider 2B at 77.1 with about a quarter of the parameters, and posts the top Civil Comments toxicity score of 95.8.

MarkTechPost gives model sizes of 3.12 billion parameters for d1-3B and 587 million for d1-omni-600M, with context windows of 32,768 and 16,384 tokens. It also points out that Liquid ran the official Decision Index scorer itself. So these are self-scored, not leaderboard submissions.

The Decision Index split matters too. d1-3B leads the Tools category at 74.5 but trails on Knowledge at 23.8, per MarkTechPost. Fast and structured does not mean it knows everything.

Why Liquid AI d1 matters

Most AI agents today call a big generative model for every small decision, which is slow and expensive. Liquid AI d1 is built for the opposite: tiny, fast calls that slot into a pipeline. Liquid says three questions over the same input take only about 1.3 times as long as one, so batching checks is cheap.

The NVIDIA angle is real. Liquid measured latency with NVIDIA on Jetson AGX Thor, AGX Orin and Orin Nano, and showed d1-3B steering a robot simulation in Isaac Sim, served from a Jetson. That is the robotics and edge pitch in one demo.

Liquid also shared a refreshingly unglamorous training note. It built d1-3B by averaging the weights of two of its own models, then fine-tuned and merged checkpoints again. It says training on long inputs, shuffling answer options and fixing shortcuts in the data mattered more than fancier techniques. On 11 image benchmarks, MarkTechPost reports d1-3B averages 74.1, roughly level with its base model at 73.9.

How it stacks up in the decision-model race

This category got crowded fast. OpenAI launched a Decisions API, and smaller labs have shipped their own decision models and benchmarks over the past few weeks. Liquid's pitch is size and speed at the edge, plus open weights you can run yourself.

The honest framing: d1-3B looks strong for its size, but the comparisons come from Liquid's own runs, and a 12B model still beats it on the index. Independent replication will tell us more than launch charts.

What to do as a reader

If you build agents or moderation pipelines, Liquid AI d1 is worth a quick test on your own data. Start with one narrow decision, such as routing or a yes-or-no check, and compare accuracy and latency with whatever you use now.

Check the license if your company is near the $10 million revenue line, and remember d1-omni-600M is an early research checkpoint. For more on this trend, read our OpenAI Decisions API breakdown, our Clef decision models coverage and our Strands Decider 2B review.