Liquid AI released two open-weight decision models on October 7, 2026: d1-3B, a text-and-image model, and d1-omni-600M, an experimental model that accepts text with images or audio. They return structured decisions—such as yes/no answers, choices, or scores—instead of generating a natural-language reply, so they are aimed at classification, routing, ranking, moderation, and similar tasks rather than general chat.
What are d1-3B and d1-omni-600M?
Liquid AI describes the d1 family as decision models: a caller provides a state (for example, text or JSON) and one or more named questions, and the model returns typed answers or probabilities in a forward pass. “Zero output tokens” means the model does not generate a textual completion; it does not mean that the request has no input, computation, latency, or response data.
The release date matters. Liquid AI’s October 5, 2026 post announced its hosted d1 API and said open weights for upcoming models were planned. The open-weight release of these two models followed on October 7.
How the two models differ
| Specification | d1-3B | d1-omni-600M |
|---|---|---|
| Size and foundation | 3.12 billion parameters; built on LFM2.5-VL-3B. | 587 million parameters; built on LFM2.5-Encoder-350M with separate vision and audio encoders. |
| Inputs | Text, JSON, images, or a mix of text and images. | Text with images, or text with audio; the model card specifies audio clips up to 30 seconds. |
| Outputs | Typed yes/no, choice, or score answers; no generated output tokens. | Typed decision answers; no generated output tokens. Not a chat model. |
| Maturity | Open-weight release. | Experimental/early research model under active development. |
| Published inference figures | Hardware-specific measurements appear in its model card. | No inference figures were reported with the release. |
Sources: d1-3B model card and d1-omni-600M model card.
#1 Best Overall
What can decision models do?
These models are best understood as components inside an application, not as replacements for a conversational assistant. A service can send a state and defined questions, then use the returned values to make a downstream choice. Potential fits include:
- Classifying or routing incoming text and structured records.
- Scoring or ranking candidates against a defined criterion.
- Moderation, checks, and guardrails that return a typed judgment.
- Visual inspection tasks for supported image inputs.
The task still needs a suitable question definition and an application that interprets the answer. A decision result is not, by itself, an explanation or a user-facing response.
Can d1-3B process images, and can d1-omni-600M understand audio?
Yes. d1-3B accepts text and images, including mixed text-and-image inputs. d1-omni-600M supports text with images or text with audio; its model card specifies audio clips up to 30 seconds. The release does not provide vision or audio decision benchmark scores: Liquid AI says Decision Index v0.3 has only a private vision split and that audio decision benchmarks remain an open problem.
The d1-3B model card separately reports 74.1 across 11 public image benchmarks, compared with 73.9 for its LFM2.5-VL-3B base. That is an image-benchmark result, not a d1 decision-benchmark score.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #2
What do the reported benchmark results show?
Liquid AI reports d1-3B at 48.57 on Decision Index 0.2.1 and describes it as the top model under 10 billion parameters in that comparison. Its October 7 release article reports a seven-dataset mean of 82.9 for d1-3B and 78.4 for d1-omni-600M. These are company-reported results; no independent evaluation is established here.
There is a version discrepancy worth keeping in mind when looking at individual d1-3B tasks. The release article reports SQuAD 2.0 at 83.3 and BoolQ at 86.3, while the currently displayed d1-3B model card reports 85.3 and 86.7. Both sources show a seven-task mean of 82.9. The omni model card retains the 78.4 mean. Do not combine individual rows from the two d1-3B tables as if they came from one consistent table.
Sources: Liquid AI’s October 7 release article and the current d1-3B model card.
How fast does d1-3B run?
Liquid AI’s d1-3B model card reports warm, one-request, single-question measurements. These are device- and request-specific, not a universal latency guarantee.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #3
| Device | Reported time | Measurement context |
|---|---|---|
| NVIDIA RTX 4090 | 8 ms | Warm, one request, one question. |
| AMD MI325X | 9 ms | Warm, one request, one question. |
| Apple M5 Pro | 30 ms | Warm, one request, one question. |
| NVIDIA Jetson AGX Thor | 16 ms | Warm, one request, one question. |
| Jetson AGX Orin 64 GB | 26 ms | Warm, one request, one question. |
| Jetson Orin Nano | 50 ms | Warm, one request, one question. |
Input size can change the result substantially. On Jetson Orin Nano, the card reports 1,640 ms for a 3.4K-token state and 202 ms for a 384-pixel image. These figures come from the model card’s stated benchmark configuration; they should not be read as comparable to every application’s workload. No inference numbers were reported for d1-omni-600M.
Source: Liquid AI d1-3B model card.
How can you run the models?
Use open weights locally
The release and model cards provide Transformers-based loading examples and describe serving routes such as vLLM and SGLang. The examples use trust_remote_code=True; review the repository code and the implications of executing remote model code in your environment before using that setting. The d1-3B card also points to Docker Model Runner and quantization discovery paths.
Start with the specific model card for its current instructions: d1-3B or d1-omni-600M. Confirm hardware, software requirements, and repository changes against the card before deployment.
Use a hosted API
Liquid AI’s October 5, 2026 announcement documents its hosted d1 API and says billing is based on input tokens only; it also gives an image-token accounting example. That post named Vercel and OpenRouter availability for text-only d1 at the time, with vision described as forthcoming on those providers. This is dated availability information, not a guarantee of current service coverage.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Source: Liquid AI’s October 5 d1 API announcement.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What does the license allow?
Both Hugging Face model cards display the license label lfm1.0. That label alone does not establish the full conditions or commercial permissions. Read the linked license text for the specific model and intended use before adopting it.
Which model should you consider?
- Consider d1-3B when you need a released text-and-image decision model and can accommodate its larger 3.12-billion-parameter footprint.
- Consider d1-omni-600M when its text-plus-audio path is relevant and you can accept an explicitly experimental model with no published inference figures.
- Consider the hosted API when you want a managed service rather than local model deployment; check current supported modalities and billing details directly with the service.
The two model cards and release materials are current as of their publication and may change; verify model availability, license terms, and implementation instructions before relying on them.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
