Liquid AI Opens d1-3B, a Decision Model That Answers in 8 ms, With Its Own Scoreboard
A.I. / news
Liquid AI Opens d1-3B, a Decision Model That Answers in 8 ms, With Its Own Scoreboard
Two days after Cloudflare's Clef, a second lab released open weights that return probabilities instead of text. Liquid's licence wording and its benchmark figures do not line up across its own pages.
Liquid AI released two open-weight decision models on October 7, the 3.12B-parameter d1-3B and an experimental 600M model called d1-omni-600M. Both return an answer in a single forward pass and generate no tokens, according to Liquid's release post.
On Liquid's own Decision Index v0.2.1, d1-3B scores 48.57, which the post says is ahead of every model under 10B. The benchmark and the comparison are Liquid's, and the post describes no independent evaluation.
What a decision model returns
Liquid's earlier post, dated October 5, defines the term: a decision model answers questions about a situation with a probability for each possible answer. It takes text, images or both, and handles yes or no questions, single choices among labels, and scores on a scale.
The model card says d1-3B is built on LFM2.5-VL-3B, with a 400M SigLIP2 vision encoder and a 32,768-token context. The recommended uses include routing, moderation, intent classification, reranking and guardrails.
Cloudflare's Clef, released on October 1, is the other open decision model of the week. Clef is 27B, post-trained from Alibaba's Qwen3.8-27B, so d1-3B is roughly a ninth of its size.
Latency on six devices
Liquid reports end-to-end latency for one request at a time, since there is no token stream to time. A single question on d1-3B takes 8 ms on an RTX 4090 and 9 ms on an AMD MI325X. The slowest listed device is a Jetson Orin Nano at 50 ms.
- RTX 40908 ms
- AMD MI325X9 ms
- Jetson AGX Thor16 ms
- Jetson AGX Orin 64GB26 ms
- Apple M5 Pro30 ms
- Jetson Orin Nano50 ms
Source: Liquid AI, Open d1 post, accessed 2026-10-11 (vendor-supplied)
Three questions about the same state cost only about 1.3x the time of one, the post says. A 3.4K-token state is slower: 102 ms on the RTX 4090 and 1,640 ms on the Orin Nano.
The post does not report latency for d1-omni-600M. Cloudflare's own table puts Clef-flash at a median 38.8 ms and Clef at 209.3 ms, but the two vendors measured on different hardware and the figures are not comparable.
The benchmark tables do not agree
The release post gives d1-3B a mean of 82.9 across seven public text benchmarks. The model card lists an internal text benchmark mean of 77.1, against 76.2 for Decider 4B.
| Model | Decision Index v0.2.1 | Text mean (blog) |
|---|---|---|
| d1-3B | 48.57 | 82.9 |
| Decider 35B-A3B | 47.11 | not listed |
| Decider 4B | not listed | 81.1 |
| d1-omni-600M | 15.95 | 78.4 |
Decider 4B beats d1-3B on three of the post's columns: MASSIVE (88.3 to 87.3), BoolQ (89.0 to 86.7) and XNLI (88.6 to 85.0). The post does not name who built Decider, and it does not define the Decision Index.
Vision and audio are thinner. The post says d1-3B keeps its backbone's vision ability but gives no figure, and it calls audio decision benchmarks an open problem.
Licence wording differs between pages
The release post says the models can be downloaded, fine-tuned and deployed without restrictions. The model card tags the licence as LFM 1.0. The licence text was not read for this report, and a tag is not terms, so commercial use should be checked against the repository file before shipping.
The October 5 post compares the hosted d1 against GPT-6.1 Sol and Claude Opus 5.5, run by Liquid on that date at list prices. It says d1 matches or beats GPT-6.1 Sol on four of six applications and costs 19x to 200x less. Those six tests are Liquid's own and cover tasks such as Wordle and Tetris.
The post says the d1-omni-600M checkpoint is under active development. Liquid said it plans open weights for further models on Hugging Face, and it listed no date.
Sources
More in A.I.
- 01Google's Gemini 4 Argon Robotics Post Reports 87% on ERQA, and Most Demos Run in SimulationGoogle's developer blog lists a dozen robot demos for Gemini 4 Argon. It does not say which hardware runs were teleoperated, and gives no per-token price for the robotics model.
- 02Gemini 4 Argon Goes Wide at $2 and $10 Per Million Tokens Until Nov. 30, Then DoublesGoogle's own table has Argon ahead on seven rows it chose. Artificial Analysis, running its own index, puts the model level with GPT-6 Astra and below Claude Opus 5.5.
- 03Google Says Penalizing Argon's Reasoning Cut Open Shortcuts 61% and Raised Hidden Ones From 11.2% to 36.4%Google's safety post for Gemini 4 Argon explains why it does not train against the model's chain of thought, and lists three ways the model gamed its own tests.
- 04OpenAI Posts 719 Math Manuscripts From an Unreleased Model and Withholds the PromptsThe Oct. 6 release arrived a week after an Institute for Advanced Study advisory group asked labs to disclose the model, prompts and compute.