Liquid d1
Liquid AI's open decision models: give them a state and questions, get calibrated answers in one pass without generated tokens.
In short. d1 is a family of small models from Liquid AI that make quick decisions instead of writing text. You describe a situation in words, a picture, a video or a sound, and ask questions such as whether it is a refund request or which team should handle it. The model answers each question with a probability for every option in a few milliseconds. It is free for commercial use until your company reaches US$10 million in annual revenue.
Summary
Liquid AI released d1-3B and d1-omni-600M on 7 October 2026, and d1-2B and d1-Omni-2B, built on its LFM3 backbone, appeared on Hugging Face in late October. They are what the field calls decision models. They do not generate tokens. Instead you give a state, which can be text, JSON, pictures, a video or audio, plus a set of named questions, and the answers are read from a single forward pass. Question types include yes or no, one or several choices from named options and ordered scores, and d1-2B also extracts text, numbers and dates, draws boxes around objects and estimates quantities with an interval. d1-3B is trained from LFM2.5-VL-3B, d1-omni-600M from the LFM2.5-Encoder-350M encoder, d1-2B from LFM3-VL-2B and d1-Omni-2B from LFM3-Omni-2B, which adds audio.
The numbers are Liquid's own. On the public split of its Decision Index 0.2.1, d1-3B scored 48.57, which Liquid says is ahead of every model under 10B and above the 36B Decider 35B-A3B at 47.11, although a 12B model in the same table, Winnow-12B, scored 50.02. On the newer 0.3 suite, d1-2B scored 51.1 against 49.0 for both d1-3B and d1-Omni-2B and 18.3 for the experimental d1-omni-600M. Liquid measured d1-3B answering one question in 8 ms on an RTX 4090, 30 ms on an Apple M5 Pro and 50 ms on a Jetson Orin Nano, and d1-2B in 5.7 ms on an H100. The models run in Transformers with remote code, and Liquid says llama.cpp support was there on day one.
The weights use the LFM Open License v1.0, which is free commercially below US$10 million of annual revenue. The cards are clear about limits. These are not chat models and write no text, knowledge-heavy questions favor larger models, and d1-2B answers get less reliable beyond about 64K tokens of state. The audio of d1-omni-600M was trained on English requests between a speaker and an assistant and clips are cut at 30 seconds. In the 30 days to 11 November 2026, Hugging Face counted about 27,900 downloads of d1-3B, 11,800 of d1-omni-600M, 11,300 of d1-2B and 4,100 of d1-Omni-2B.
Good for
- Routing, triage, moderation and guardrail steps inside agents, where you want a probability to set a threshold on
- Reading text, numbers and dates from documents and locating objects in pictures with d1-2B
- Edge devices such as Jetson boards and laptops, since d1-3B answers a single question in 50 ms or less on every device Liquid measured
- Routing voice messages and sound events with d1-Omni-2B
Watch out for
- It cannot chat, explain itself or call tools, so use a chat model where you need text
- Knowledge-heavy questions favor larger models, and long states lose reliability
- It needs Transformers with remote code, version 5.14 or later for d1-3B and 5.19 or later for d1-2B and d1-Omni-2B, or llama.cpp
- The benchmarks and latencies are Liquid's own, and commercial use stops being licensed at US$10 million of annual revenue
- d1-omni-600M is an early research release, and its audio was trained only on English requests
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the d1-omni-600M variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| d1-omni-600M | 587M | 1.35 GB | 0.675 GB | 0.405 GB | Laptop or phone |
| d1-2B BF16 | 1.92B | 4.42 GB | 2.21 GB | 1.32 GB | Laptop or phone |
| d1-Omni-2B BF16 | 2.17B | 4.99 GB | 2.5 GB | 1.5 GB | Laptop or phone |
| d1-3B BF16 | 3.12B | 7.18 GB | 3.59 GB | 2.15 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
An Apache-2.0-style license from Liquid AI with one big catch. Commercial use is licensed only while your company, together with its affiliates, earns under US$10 million a year. Above that line you are not licensed and need your own agreement.
- Free commercial use only below a revenue cap
See also
Frequently asked questions
- What is Liquid d1?
- d1 is a family of small models from Liquid AI that make quick decisions instead of writing text. You describe a situation in words, a picture, a video or a sound, and ask questions such as whether it is a refund request or which team should handle it. The model answers each question with a probability for every option in a few milliseconds. It is free for commercial use until your company reaches US$10 million in annual revenue.
- Can I use Liquid d1 commercially?
- Yes, with conditions. Liquid d1 is released under the LFM Open 1.0, which allows commercial use but adds extra rules: Free commercial use only below a revenue cap. Read the license text before you ship a product.
- Is Liquid d1 open source?
- Liquid d1 is open-weight, not open source in the strict sense. You can download the model files, but the LFM Open 1.0 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Liquid d1 on my own computer?
- Yes. The smallest variant, d1-omni-600M, needs about 0.405 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Liquid d1 released?
- Liquid d1 was released by Liquid AI on October 7, 2026.
- How long a context window does Liquid d1 have?
- Liquid d1 supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Liquid AI: Open d1, edge decision models for text, vision and audio (7 Oct 2026) liquid.ai
- d1-3B model card (Hugging Face) huggingface.co
- d1-2B model card (Hugging Face) huggingface.co
- d1-Omni-2B model card (Hugging Face) huggingface.co
- d1-omni-600M model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Clef Cloudflare | 9.41B – 27.4B | 64K | Apache 2.0 (Permissive) | 6.49 GB Laptop or phone | |
| AutoTrust Decider AutoTrust AI Lab | 8.95B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 6.18 GB Laptop or phone | |
| LFM3 Liquid AI | 1.73B – 2.17B | 256K | LFM Open 1.0 (Conditional) | 1.19 GB Laptop or phone | |
| LFM3-Audio-2B Liquid AI | 2B | 32K | LFM Open 1.0 (Conditional) | 4.6 GB Laptop or phone |