Integrate d1 as a decision step inside your existing agent: send it the current state and a bounded question, interpret the returned probabilities in your application, validate and execute an allowed action, then collect the next state. d1 returns structured probabilities for decisions; Liquid’s launch material does not present it as a complete tool-executing agent framework.
How d1 fits into an agent
Liquid AI describes d1 as a decision model that accepts text, images, or both, plus one or more questions, and returns probabilities without generating tokens. Its documented question types cover whether something is true (noul), which label to select (choice), and how to rate a state across a scale (score). These are useful when an agent’s next decision can be expressed as a defined set of outcomes.
The surrounding application remains responsible for the agent loop: gathering state, deciding how to interpret probabilities, enforcing policy, executing tools, persisting state, and handling retries and failures. Liquid’s agent guidance likewise describes an agent as the combination of a model and its harness. See Liquid AI’s agentic AI overview and its d1 launch post.
Call the Liquid AI API
Liquid AI’s October 5, 2026 launch post documents d1 on the Liquid AI API under model name d1. It directs developers to create an API key in the Liquid AI Console at Dashboard → API Keys. The post’s example uses a bearer token and the https://api.liquid.ai/decisions/v1/systemone endpoint:
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute#1 Best Overall
response = requests.post(
"https://api.liquid.ai/decisions/v1/systemone",
headers={"Authorization": f"Bearer {LIQUID_API_KEY}"},
json={
"model": "d1",
"state": "Camera image of a circuit board on the production line.",
"questions": {
"defect": {
"type": "noul",
"instructions": "Does this circuit board have a defect?"
}
}
},
)The launch example reads the answer probability at response.json()["answers"]["defect"]["noul"]. Treat this as an illustration of the launch post’s request and response handling, not as a full production specification: consult the official d1 documentation linked from Liquid’s announcement for the current schema and validation requirements.
Choose the question type for the decision
Use noul for a binary condition
Ask a yes-or-no question such as whether a visible part is defective or whether a support ticket meets an escalation condition. The result is a probability between 0 and 1. Your application must define what probability, if any, triggers a particular action; the launch post does not prescribe a universal threshold.
Use choice to select among available outcomes
Provide the actual labels or options available to the agent, such as permitted navigation actions on a page. Liquid’s launch example shows a web agent selecting its next action from options on a flight-search site. Validate the returned selection against the application’s allowed actions before invoking a tool.
Use score to rate a state
Use a scale when the downstream decision depends on an ordered rating rather than a binary result or a single label. Liquid describes the output as weighted probabilities over scale levels, so the application should decide how those probabilities map to its own policy.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Ask more than one question about the same state when useful
Liquid says a request can include multiple questions about the same state. Each question is billed as its own prompt, however, so combine questions only when the additional decisions are worth their added usage.
Build the agent loop around the decision response
- Gather the current state. Collect the text, page details, tool output, or image that the next decision needs.
- Ask a bounded question. Use
noul,choice, orscoreand describe the decision and available outcomes clearly. - Apply application logic. Interpret the probabilities using your own thresholds, fallback rules, or selection policy. Do not treat a probability as authorization to act.
- Validate and execute. Check any selected action against the tools and arguments your application permits, then execute it in the surrounding harness.
- Collect the next state. Feed the result of the action back into the loop for the next decision.
This pattern uses d1 for the decision it documents while leaving tool execution and control flow in the application. For tasks that need free-form explanations, generated content, or a model to perform multi-step tool orchestration itself, d1’s probability-returning interface may not be sufficient by itself.
Send screenshots or other images
Liquid’s October 2026 example encodes a JPEG image as a base64 data URL and sends it in the images array alongside a textual state and named question:
import base64
import requests
with open("screen.jpg", "rb") as image_file:
encoded = base64.b64encode(image_file.read()).decode("ascii")
response = requests.post(
"https://api.liquid.ai/decisions/v1/systemone",
headers={"Authorization": f"Bearer {LIQUID_API_KEY}"},
json={
"model": "d1",
"images": [f"data:image/jpeg;base64,{encoded}"],
"state": "A flight-search page is open. Choose the next permitted action.",
"questions": {
"next_action": {
"type": "choice",
"instructions": "Which available action should the agent take next?"
}
}
},
)The image example is the launch post’s documented format, not proof that every image type, size, or encoding is accepted. Confirm supported formats and limits in the live API reference before relying on a particular screenshot pipeline.
Account for usage, availability, and performance claims
Liquid’s October 5, 2026 launch post states that d1 is billed on input tokens only, at $0.04 per million input tokens, with each question charged as its own prompt and images counted at the same input-token rate as text. It says images count as 1.5 input tokens per 32×32-pixel patch; the post’s example counts a 1024×1024 image as 1,536 tokens. These are launch-post figures, not a substitute for checking current pricing, plan conditions, or request limits.
The same post says d1 was available through Liquid AI’s API and through Vercel and OpenRouter at launch; it described text-only support through the latter providers, with vision forthcoming. Provider availability can change, so verify present support before choosing an integration route.
Liquid reports 200–300 ms for text decisions. Its launch post also reports 85–97% accuracy across four production lines using the public VisA dataset, and says d1 solved Wordle from screenshots without a manually constructed textual board. These are vendor-reported demonstrations, not independent benchmarks or guarantees for a different workload.
Liquid further reports that, in a comparison run once per application on October 5, 2026, d1 matched or beat GPT-6.1 Sol on four of six applications and cost 19× to 200× less in that comparison. The post says listed prices and up to eight requests in flight were used. It also reports removing 52% of tokens in a coding-agent context-compaction application while retaining outputs needed for the task. Treat all of these as dated vendor comparisons and demonstrations, not general performance expectations.
Recommended Free Tools
Best Value
Decide whether d1 is the right call for a step
d1 is a natural option when the agent needs to classify, route, or score a state against fixed outcomes and probability values are useful to downstream logic. Compare it with a general language-model call based on these factors:
- Outcome shape: Can you define the possible answers in advance, or does the task need an open-ended response?
- Probability handling: Will your application use probabilities to set thresholds, choose among options, or route work?
- Input modality: Does the decision need text alone or visual input as well?
- Latency and cost: Measure the current service behavior and cost for your own request volume and input sizes rather than assuming launch figures apply.
- Need for explanation or generation: d1 is described as returning probabilities without token generation, so a task that requires written reasoning or generated user-facing content may need another component.
Liquid’s separate LFM2.5-2.6B is not a local version of d1. Liquid’s August 4, 2026 release describes LFM2.5-2.6B as an on-device model trained for agentic workloads such as planning, tool use, and multi-step tasks, with weights available on Hugging Face. It is a distinct model and deployment path; see Liquid AI’s LFM2.5-2.6B release post.
Check the production details before shipping
Liquid’s launch post is not a complete production API reference. Before deploying, confirm the live documentation’s request schema, supported image formats and limits, rate limits, error and timeout behavior, retry guidance, and service terms. The launch material does not establish these details, so avoid assuming a particular failure-handling contract or request ceiling.
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.
