Liquid AI publishes open-weight d1 decision models for the edge

What did Liquid AI release, and when?

The Hugging Face blog post “Multimodal open d1 decision models for the edge,” dated 7 October 2026 and signed Aurelien Lac and fernandofernandes, says Liquid AI is releasing two open decision models: d1-3B and d1-omni-600M (experimental). The d1-3B card lists 3.12 billion parameters and a SigLIP2 NaFlex 400M vision encoder. The d1-omni-600M card lists 587 million parameters: a 381 million shared trunk and decision head, a 94 million vision encoder and a 112 million audio encoder. Both Hub repositories were created on 5 October.

A separate Liquid AI blog dated 5 October introduced a hosted d1 API with vision, billed at $0.04 per million input tokens, and said open weights were planned. That post is still live. The 7 October Hugging Face post is the weight drop. The citation on both cards points to https://www.liquid.ai/blog/open-d1; that URL returned 404 on 7 October, so this article uses the Hugging Face post and the two cards.

How do the open d1 models work?

Decision models in this family do not write text. A caller supplies a state—text, JSON, images, or, on d1-omni-600M, a 16 kHz mono clip—and a set of named questions. Each question is a noul (yes/no probability), a choice among named options, or a score over two to ten ordered levels. The model reads the state once and returns a probability for every allowed answer, with usage.output_tokens reported as 0. That is the same three-way shape as OpenAI’s Decisions API, which is a hosted gpt-6-luna endpoint, not a downloadable weight.

d1-3B is trained from the decoder-only LFM2.5-VL-3B vision-language model and accepts text and images, with a 32,768-token context. d1-omni-600M is trained from the bidirectional LFM2.5-Encoder-350M trunk. It accepts text with images or text with audio, not both in one request; passing both raises ValueError. Its context is 16,384 positions. With images, the card says state and question text is cut to 896 tokens, as trained. Audio training, the card says, used English speaker-and-assistant clips and cuts audio at 30 seconds.

The cards recommend routing, triage, moderation, intent classification, extraction checks, reranking, judge scoring, agent guardrails and visual inspection. They also say the models are not chat models and do not write text. The 5 October hosted-d1 post described playground demos such as filing tickets and inspecting public VisA parts; those demos are vendor examples, not tests we ran.

What scores and speeds does Liquid AI report?

The 7 October post and both cards say Liquid AI scored d1-3B and d1-omni-600M with the official Decision Index 0.2.1 scorer and did not submit those rows to the public leaderboard. Other rows come from that public board. Winnow-12B leads at 50.02. d1-3B is next at 48.57, ahead of Decider 35B-A3B at 47.11 and of every listed 4B and 9B model. d1-omni-600M is last at 15.95. The post’s “best decision model under 10B” line is that vendor ranking, not an independent bake-off.

The d1-3B card’s “benchmarks as decisions” table reports a mean of 77.1 across SQuAD 2.0 (85.3), Civil Comments, MASSIVE intent, HelpSteer2, PubMedQA, BoolQ, XNLI and PAWS-X, against 76.2 for Decider 4B. The Hugging Face blog table omits HelpSteer2 and lists a mean of 82.9 with SQuAD 2.0 at 83.3. Those two Liquid AI tables do not match. We repeat both and do not pick a winner. The card also reports 74.1 on eleven public image benchmarks read as decisions, against 73.9 for the LFM2.5-VL-3B base; with images removed, the same questions score 45.1. d1-omni-600M’s matching text-only mean in the blog is 78.4. The post says Decision Index v0.3 has only a private vision split and that audio decision benchmarks are an open problem, so it reports no vision or audio Decision Index numbers.

How do you run the weights, and what licence applies?

The d1-3B card asks for transformers 5.14 or newer, torch, torchvision and pillow, and AutoModel.from_pretrained("LiquidAI/d1-3B", trust_remote_code=True). d1-omni-600M asks for transformers 5.15 or newer plus soundfile, and the card warns against bfloat16 on that model. Custom code on a Hub card is a load-time trust decision; our note on a backdoored coding model in Codex CLI is about a different incident, but it is the right class of caution. We did not execute the shipped modeling_d1.py.

Both cards set license: other and license_name: lfm1.0. The LICENSE file is “LFM Open License v1.0.” Section 5 conditions commercial use on the licensee’s legal entity staying under a $10 million annual-revenue threshold. Entities at or above that threshold are not licensed for commercial use. Qualified non-profits using the work for non-commercial or research purposes are carved out of that threshold. That is not Apache 2.0. For the difference between a downloadable weight and a permissive licence, see open weights versus open source.

The post points to Hugging Face downloads and a System One Arcade Space for demos. The hosted 5 October API remains a separate product with its own price. The cards say d1-3B answers a single question in 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin 64 GB, 50 ms on a Jetson Orin Nano, 8 ms on an RTX 4090 (with torch.compile reduce-overhead; 16 ms without) and 9 ms on an AMD MI325X. Those are vendor warm-call medians. The post reports no speed numbers for d1-omni-600M.

How does this sit next to other decision models?

OpenAI’s Decisions API, opened as a public beta on 6 October, uses the same predicate/choice/score idea on gpt-6-luna and bills $0.10 per million input tokens. You cannot download that model. Liquid AI’s 5 October hosted d1 listed $0.04 per million input tokens. The 7 October weights are a local option for teams that can accept LFM 1.0 and custom Hub code.

The cards also name other open deciders on the 0.2.1 board. Those rows are copied from the public leaderboard, not rerun here. A typed decision is still a classifier: you need labels and a threshold before you trust a 0.92 noul. That is the same rule as evaluating an AI agent on the task you actually run, not on a vendor table.

What should readers not assume?

Open d1 is not a general assistant, not an Apache dump, and not a measured replacement for gpt-6-luna Decisions. The 48.57 figure is Liquid AI’s unofficial 0.2.1 score. The two Liquid AI public-benchmark tables disagree on SQuAD 2.0 and on the mean. d1-omni-600M’s 15.95 index score and missing latency numbers are part of the release, not a footnote. We have not timed a Jetson, inspected a production ticket queue, or confirmed that the Arcade Space matches the cards.