Quick Answer
Liquid AI’s d1-3B is an open-weight decision model designed for fast, structured classifications rather than chat-style text generation. Released October 7, 2026, it accepts text and images and returns an answer in one forward pass. Its reported speed and small size suit local routing and moderation workloads, but Liquid AI’s performance figures are self-reported. Review the model card before deployment.
Key Takeaways
- Liquid AI released d1-3B and d1-omni-600M as downloadable open-weight models on October 7, 2026.
- Liquid AI d1-3B uses 3.12 billion parameters and accepts text plus image inputs.
- Liquid AI’s decision models provide structured answers without generating output tokens.
- Liquid AI reports 8-millisecond d1-3B latency on an Nvidia RTX 4090 for one decision.
- Liquid AI d1-3B is designed for routing, triage, moderation, scoring, and other bounded decision tasks.
What did Liquid AI release with d1-3B?
Liquid AI released two downloadable open-weight decision models, d1-3B and d1-omni-600M, on October 7, 2026. Liquid AI describes the releases as models for structured decision tasks, where a system must classify, route, score, or approve an input instead of producing a conversational response. Liquid AI’s launch announcement identifies both models as part of its decision-model lineup.
Liquid AI d1-3B is the larger release and is aimed at text-and-image decisions. Liquid AI d1-omni-600M is the smaller multimodal release, with support for text plus images or text plus audio. The two models are related, but they are not interchangeable because their supported inputs, size, and development status differ.
The most important distinction is that Liquid AI positions these models as alternatives to generative systems for narrow workflows. A generative model normally writes a sequence of output tokens. A decision model can instead return a structured result such as a category, score, route, or safety determination. That approach is most useful when the application needs a consistent operational result rather than a written explanation.
Liquid AI d1-3B also arrives as interest grows in models that can run closer to the device or workstation handling the data. Users considering local AI should still separate local execution from automatic privacy guarantees. A local model can reduce the need to send an input to a remote service, but the full privacy outcome still depends on the app, logging configuration, connected services, and the data supplied to the model.
How does Liquid AI d1-3B make decisions without generating text?
Liquid AI d1-3B makes a structured decision in one forward pass, with zero generated output tokens, according to Liquid AI. A forward pass is the model’s computation over the supplied input. Instead of repeatedly predicting the next word or token, the model produces the requested decision result directly. Liquid AI’s decision-model documentation describes this design for classification and routing workloads.
Liquid AI’s approach matters because token-by-token generation can add delay when an application only needs a limited answer. A support system, for example, may only need to send a ticket to the correct team. A moderation system may only need to apply a defined label. In those cases, generated prose can be unnecessary work rather than a useful product feature.
Liquid AI d1-3B does not replace a general chatbot for every task. A reader who needs a detailed summary, a draft email, code, or a multi-step explanation needs a model that generates language. Liquid AI d1-3B is a closer fit for applications where the desired output is predefined and can be validated by the software around it.
Developers building systems that combine routing with generative AI can use a decision model as an earlier filter. That design can direct simple requests to a lower-cost path and reserve a language model for tasks that need a written answer. The same kind of operational separation matters as AI systems become more autonomous, including systems such as work agents that run for extended periods, where clear guardrails can limit what an agent is allowed to do.
What are the Liquid AI d1-3B specifications?
Liquid AI d1-3B has 3.12 billion parameters, accepts text and images, and supports a 32,768-token context window. The context window is the maximum amount of input material the model can consider in one request. The Liquid AI d1-3B model card on Hugging Face provides the specifications and download access.
Liquid AI d1-3B’s text-and-image support is relevant for workflows that need to assess a document, screenshot, product image, or visual inspection input alongside written instructions. A visual inspection system could use an image plus a structured classification request. A support workflow could use a screenshot plus a routing label. The supported inputs do not mean the model automatically understands every specialized image or business policy, so developers need task-specific evaluation before relying on a result.
| Model | Parameter count | Supported inputs | Context window | Important limitation |
|---|---|---|---|---|
| Liquid AI d1-3B | 3.12 billion | Text and images | 32,768 tokens | Designed for structured decisions, not free-form generated responses |
| Liquid AI d1-omni-600M | 587 million | Text plus images, or text plus audio | 16,384 tokens | Cannot accept image and audio inputs together in one request |
Liquid AI d1-omni-600M supports up to 30 seconds of speech input, which gives it a separate role in voice-command routing and similar short audio tasks. Liquid AI describes d1-omni-600M as an early research release under active development. The practical response is to treat the smaller model as a model to test carefully, especially when a decision affects access, safety, money, employment, health, or another sensitive outcome.
How fast is Liquid AI d1-3B in reported testing?
Liquid AI reports that d1-3B makes one decision in 8 milliseconds on an Nvidia RTX 4090, 9 milliseconds on an AMD MI325X, and 30 milliseconds on an Apple M5 Pro. Those figures are self-reported and describe a single decision, so they should not be treated as a guarantee of end-to-end application speed. Hardware configuration, input preparation, software integration, and the surrounding workflow can all affect a deployed system’s response time.
Liquid AI d1-3B’s reported latency is important because a decision model may sit in front of another system. A content platform can evaluate an upload before publication. A customer-service application can route an incoming request before assigning it to a person or a larger model. A voice interface can determine which command path to use before responding to the user.
Liquid AI also says d1-3B scored 48.57 on the public split of Decision Index v0.2.1 and ranked ahead of every model below 10 billion parameters in its comparison. The benchmark result is self-reported, and it evaluates a particular public benchmark split rather than every possible routing or classification task. Organizations should test the model on representative data and measure incorrect decisions, not only speed.
For most users, the relevant question is not whether 8 milliseconds is faster than a chatbot in isolation. The relevant question is whether a rapid structured decision improves the product without creating an unacceptable error path. A wrong support route can be corrected. A wrong safety or financial decision may require human review before the application acts on it.
What work is Liquid AI d1-3B designed to handle?
Liquid AI documents d1-3B and related decision models for content moderation, support-ticket routing, triage, visual inspection, voice-command routing, agent guardrails, and scoring. These are bounded tasks with a defined output space, such as selecting one category from an approved set or assigning a score under a specified policy.
Content moderation is a useful example because the model can assign a label that determines whether another system escalates, blocks, or queues material for review. The limitation is that a structured output can still be wrong, especially when context, language, irony, or policy interpretation matters. The safest design keeps a human escalation path for uncertain or high-impact decisions.
Agent guardrails are another relevant use because a decision model can assess whether an agent request fits a defined rule before the agent takes an action. That does not make an agent safe by itself. Guardrails need clear policies, access controls, logging, and limits on the actions the software can perform. Readers evaluating AI-generated content should also understand that tools such as AI content detection systems can have defined technical scope and do not establish the truth or safety of every output.
The practical value of Liquid AI d1-3B is therefore operational consistency, not a claim that automated classification is inherently correct. Organizations should define which decisions can be automated, which decisions need review, and what happens when the model produces a low-confidence or disputed result.
Can Liquid AI d1-3B run locally?
Liquid AI d1-3B can be downloaded from Hugging Face, and Liquid AI says the model has day-one llama.cpp support. Local availability can help developers test or deploy the model on compatible hardware without relying on a hosted chat interface. The actual deployment path still depends on the hardware, inference software, model format, and the application’s engineering requirements.
Liquid AI d1-3B’s reported Apple M5 Pro latency indicates that Liquid AI tested the model on Apple hardware, while the Nvidia RTX 4090 and AMD MI325X figures show testing on separate high-performance platforms. Those results do not establish that every laptop, desktop, or phone will deliver similar performance. A model’s parameter count is only one part of local deployment planning.
Local AI can be attractive for teams that need tighter control over inputs, especially where documents, images, or support requests contain sensitive information. A local setup still needs ordinary security controls. Restrict access to model inputs, protect stored logs, review connected tools, and remove unnecessary data before testing. Hardware-focused local systems are also becoming a distinct category, as shown by the growing market for AI computers that keep data at home.
Developers should stop before treating local installation as a substitute for a security review. Contact the software vendor or an internal security team when the model will process regulated data, make high-impact decisions, connect to production systems, or receive untrusted files from the public.
What limits should developers consider before using Liquid AI d1-3B?
Liquid AI d1-3B’s main limit is that it is built for structured decisions rather than general text generation. The model can be useful when the application has a clear answer format, but it is not the appropriate tool for every language task. Developers need to define the possible outputs, validate them in software, and confirm that the model is evaluated on the actual inputs it will receive.
Liquid AI’s published benchmark and latency figures are useful starting points, but they remain vendor-reported results. Independent testing may produce different outcomes, particularly when the model receives longer inputs, varied image types, specialized vocabulary, or difficult edge cases. The most sensible approach is to run a controlled evaluation with known examples before connecting the model to a live workflow.
High-impact uses need additional safeguards. A model should not become the sole decision-maker for medical, legal, employment, credit, safety, or emergency outcomes without appropriate expert review and accountable processes. Structured output can make automation easier to integrate, but a clean label or score does not explain whether the underlying decision was fair, accurate, or appropriate.
Liquid AI d1-3B is best understood as a specialized building block. Developers who need fast routing, moderation, scoring, or guardrails may find the architecture useful. Developers who need detailed human-readable responses should use a generative model or a combined system that gives each model a limited, auditable role.
FAQ
What is Liquid AI d1-3B?
Liquid AI d1-3B is an open-weight decision model that accepts text and images and returns structured answers without generating output tokens. Liquid AI released the 3.12-billion-parameter model on October 7, 2026.
Does Liquid AI d1-3B generate chatbot responses?
No, Liquid AI d1-3B is designed for structured decisions rather than chat-style generated text. The model is intended for outputs such as classifications, routing choices, scores, and guardrail decisions.
How fast is Liquid AI d1-3B?
Liquid AI reports 8-millisecond latency for one d1-3B decision on an Nvidia RTX 4090. Liquid AI also reports 9 milliseconds on an AMD MI325X and 30 milliseconds on an Apple M5 Pro, but real deployment speed can differ.
Can Liquid AI d1-3B analyze images?
Yes, Liquid AI d1-3B accepts text and image inputs. Image support can help with structured visual-inspection or screenshot-routing tasks, but developers should test performance on their own image data.
Is Liquid AI d1-3B suitable for high-impact decisions?
No, Liquid AI d1-3B should not be the sole decision-maker for high-impact outcomes without expert review and additional safeguards. A structured result can be useful for triage, but it does not remove the need for accountability or error handling.