Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Liquid AI released d1-ultra on October 15, 2026. It is a hosted model that accepts text, images, or both and returns probabilities for typed decisions instead of composing a prose answer. Liquid reports a first-place score of 65 on the Artificial Analysis Decision Index.
What d1-ultra is
d1-ultra is built for applications that need a measurable decision: whether an image meets a specification, which category an item belongs to, how severe a condition is, or whether a case should be escalated. Its output is a probability for one or more questions, rather than a conversational explanation.
Liquid AI describes the approach this way: “Decision models answer questions about a situation with a probability for each possible answer.” That makes d1-ultra different from a conventional language model whose primary output is generated text. An application can use the probabilities directly in a threshold, routing rule, queue, or human-review workflow.
The three request types
| Request type | What it returns | Typical use |
|---|---|---|
Noul |
A yes/no probability | Pass/fail checks, eligibility decisions, or escalation triggers |
Choice |
A selection among named labels | Classification and routing |
Score |
A position on an ordered scale | Severity, quality, or risk scoring |
Multiple questions can be evaluated against the same state. A single inspection could therefore ask separate yes/no, label-selection, and ordered-score questions about the same text and image.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
What the reported score of 65 means
Liquid AI says d1-ultra ranks first on the Artificial Analysis Decision Index with a score of 65. A contemporaneous display of that index showed d1-ultra at 65 and Jev 1.14 at 57, putting the reported gap at eight points.
This is a release-era benchmark claim, not a complete independent audit of every test. Liquid’s release does not provide a full per-test breakdown, so the composite should be treated as an indicator of performance on that index rather than proof that d1-ultra is best for every workload.
Do not confuse the Artificial Analysis composite with Liquid’s separate internal comparison of 112 evaluations. Those are different evaluation sets and should not be combined into one score.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Rank #2
- Delivers 25 TFLOPS of AI inference performance using four MemryX MX3 accelerators in a compact M.2 2280 M-key module.
- Connects through PCI Express Gen 3 and above for straightforward integration into embedded and edge systems with an M.2 M-key slot.
- Includes 42 MB of distributed on-chip weight storage to support low-latency, batch-1 pipelined inference.
- Designed for low-power deployment with a 1 W–12 W operating range and host-managed airflow requirements.
- Supports ONNX, TensorFlow, Keras, and TFLite models through the MemryX software stack on Linux and Windows hosts.
How image input and billing work
d1-ultra accepts text, images, or both. Images count as input tokens at the same rate as text. Liquid’s example uses a 1,024 by 1,024 image and a 32-by-32-patch accounting method, producing 1,536 input tokens.
Published token price
| Item | Published treatment |
|---|---|
| Input | $0.042 per million tokens through Liquid AI’s API |
| Output | No output-token charge |
| Image input | Counted as input tokens at the same rate; a 1,024×1,024 example uses 1,536 tokens |
| Question billing | Each question is billed as its own prompt, including the question text and all images |
Repeating an image across five questions therefore bills the image input five times. Using the published example, one image-question pair contains 1,536 image tokens; at $0.042 per million tokens, that image portion is about $0.0000645 before adding the question text and any other input.
Vercel’s AI Gateway documentation describes the same input-only billing model and lists a 32,768-token context window for its d1-ultra route. The actual total for a request still depends on all text and images included in each separately billed question.
Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
Where you can access d1-ultra
Liquid AI API
Liquid’s hosted API is the primary access route described with the release. It is the direct option when you want to manage request construction, question types, authentication, and application-side thresholds yourself.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Vercel AI Gateway
Vercel documents the model under the name liquid/d1-ultra. This route is useful when an application already uses the gateway’s provider abstraction, while retaining the documented 32,768-token context window and input-only billing description.
OpenRouter
OpenRouter lists a d1-ultra provider route. Liquid’s release originally described Vercel and OpenRouter support as forthcoming; Vercel has since published availability.
Rank #4
- 48GB AI graphics accelerator
Self-hosting status
The release says d1-ultra is based on a larger open-weight model than d1 and includes Liquid-specific customizations. It does not name that backbone or disclose a parameter count. No downloadable d1-ultra weights are announced in the cited release, so the documented deployment choices are hosted services rather than a confirmed local installation.
Where the model fits operational workflows
Visual inspection
For a circuit-board image, you could ask a Noul question such as whether the board meets a stated specification, then use the returned probability to pass clear cases and send borderline cases to a technician.
Recommended Free Tools
Insurance-claim checking
A customer photograph and claim details can be evaluated with a yes/no consistency question, a Choice question for claim category, and a Score question for apparent damage severity. The probabilities can feed an existing review queue instead of requiring a generated narrative for every claim.
Best Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
Routing, moderation, and guardrails
Typed decisions are also suited to classification, moderation, policy guardrails, triage, and deciding when a human must review an item. The application should define thresholds and fallback behavior; a probability is an input to a policy, not the policy itself.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.When a decision model is preferable to a prose model
| Need | Why d1-ultra’s format can help |
|---|---|
| Threshold-based automation | Probabilities can be compared directly with an application threshold. |
| Consistent labels or scales | Choice and Score constrain the answer space instead of requiring text parsing. |
| Image-based decisions | Text and visual evidence can be supplied in the same request. |
| Lower response overhead | No output tokens are charged by the published Liquid pricing, and the model need not generate an explanation for every item. |
| Need for an explanation | A conventional language model or a separate explanation step may still be necessary, because d1-ultra’s documented output is the decision probability. |
The benchmark score does not by itself establish calibration for your domain. Before automating a consequential decision, measure false positives, false negatives, and probability calibration on representative, labeled data.
Implementation checklist
- Define the decision contract: a yes/no question, named labels, or an ordered scale.
- Specify the evidence that will be sent as text, images, or both, and keep the question wording stable for evaluation.
- Choose a hosted route: Liquid’s API, Vercel AI Gateway with
liquid/d1-ultra, or the listed OpenRouter route. - Account for every question separately when estimating token cost, especially if the same image is reused.
- Set thresholds for automatic action, manual review, and rejection rather than treating every probability as an automatic answer.
- Evaluate on your own labeled examples and monitor drift, disagreement, and escalation volume after deployment.
Known limits and unanswered details
- The 65 score is a reported Artificial Analysis Decision Index result, and the release does not include a complete independent per-test breakdown.
- The index result is separate from Liquid’s internal 112-evaluation comparison.
- The model’s underlying backbone, parameter count, and downloadable d1-ultra weights are not disclosed in the cited release.
- Hosted access introduces provider, availability, and data-handling considerations that must be checked against the requirements of your application.
Bottom line
d1-ultra is aimed at systems that need probability-bearing decisions from text and images, not a chatbot-style paragraph. Liquid’s reported score of 65 makes it notable at release, while the $0.042-per-million-input-token price, separate billing for each question, and hosted-only access documented so far are the practical factors to assess before adoption.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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.
