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 →As of October 7, 2026, Liquid AI’s published d1 launch material documents access through its API, console, and playground—not downloadable d1 weights or a verified local installation path. You can try d1 through Liquid AI’s official site, but do not treat instructions for other Liquid Foundation Models as a way to run d1 locally. This guide explains what d1 returns, how to evaluate its decisions through the documented access route, and what to check if you need on-device inference.
Can you run Liquid AI d1 locally?
Not on the basis of the official d1 launch material available as of October 7, 2026. Liquid AI’s October 5 announcement documents hosted access through its API and a console/playground for trying the model. It does not provide a d1 weight download or a local installation procedure. Check Liquid AI’s official release information for changes; model availability can change.
As an Amazon Associate I earn from qualifying purchases.
Liquid AI says open weights for upcoming models are planned. That is a future-facing statement, not confirmation that d1 weights are currently downloadable. The broader Liquid Foundation Model (LFM) catalogue includes models described for local use across CPUs, GPUs, and NPUs, but a family-level capability does not establish local availability for d1. If local inference is a hard requirement, choose and verify a specific LFM model rather than assuming d1 is included. See the LFM catalogue and the individual model’s current card, license, runtime, and hardware requirements.
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1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesWhat d1 does—and what to provide it
Liquid AI describes d1 as a decision model: it takes unstructured text, images, or both, together with one or more questions, processes them in one forward pass, and returns probabilities for possible answers without generating tokens. That makes it suited to choosing among defined outcomes, rather than open-ended conversation.
#1 Best Overall
- 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.
For a decision task, prepare three things: the input material, a clear question, and the finite set of candidate outcomes. Then inspect the probabilities returned for those outcomes. The launch post establishes this general interface, but does not establish a complete request schema or API endpoint in the material cited here. Use the current console or API documentation for exact fields and authentication; do not rely on guessed code or parameters.
How to test d1’s decision-making
A useful test measures performance on decisions you actually care about, not only on an impressive demonstration. Freeze the examples, candidate labels, and reference answers before submitting them. Keep the test set unchanged while comparing results so the score has a stable meaning.
Rank #2
- ✅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
- Define the task. State what information d1 will receive, the question it must answer, and every allowed outcome. Make labels mutually understandable; if “uncertain” or “other” is a valid result, include it explicitly.
- Build a representative test set. Include ordinary cases, ambiguous cases, and edge cases from the intended use. Preserve the original input and reference answer for each item. Record the dataset, its source, and the number of examples.
- Run the frozen set. Use the documented API or playground interface. Save the input, candidate outcomes, returned probabilities, model/version information when available, and any error or refused response. Avoid changing labels or examples midway through a run.
- Score errors against the task’s risks. For binary classification, report precision, recall, and F1 when false positives and false negatives have different consequences. For multi-choice decisions, report accuracy and a confusion matrix to show which labels are mistaken for which others. If downstream decisions depend on the probability values themselves, assess calibration separately; correct rankings alone do not show that a stated probability is reliable.
- Publish enough detail to reproduce the result. Give the dataset and example count, label definitions, scoring rule, model/version, and relevant request conditions. Report mistakes as well as aggregate scores, especially where a wrong decision could cause harm.
These are evaluation recommendations, not a protocol Liquid AI has said it used for every d1 demonstration. No independent d1 replication study or fully reproducible public protocol for the launch comparison is established in the cited material.
How to interpret Liquid AI’s published examples
Liquid AI’s launch post reports several task results. They are company-reported demonstrations, not independent validation:
Rank #3
- 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
- Wordle: 12 of 12 games solved, averaging 3.8 guesses.
- Quick, Draw!: 5.2 of 6 doodles recognized; the post gives 0.6 as the result of random guessing.
- Tetris: 70 to 81 lines when supplied with a screen image, compared with describing the game in text alone.
These figures show what the company chose to demonstrate, but they do not establish accuracy across other tasks or settings. For any result you plan to rely on, run a test set that reflects your own inputs and error costs.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the d1 comparison says about cost and chat models
Liquid AI also reported d1 as 19× to 200× lower cost than two named comparison models across six applications. The company says it ran the comparison once per model on October 5, 2026, using its Playground comparison script. Its stated assumptions included a d1 input-token list price of $0.04 per million tokens, no prompt-cache discounts, and up to eight requests in flight. Some code questions and compaction sessions were written after d1’s pipeline was set. For visual inspection, each model saw a good part from the same line beside the part being inspected.
Rank #4
- Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor.
- 2.5W typical power consumption
- 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.
Read this as a vendor-reported, task-specific comparison—not a general price guarantee or independent benchmark. The result depends on the selected models, workload, pricing assumptions, and comparison setup. If you make your own comparison, keep inputs and output constraints consistent; record model versions, number of runs, concurrency, and the price assumptions used. A single run per model is not enough to characterize run-to-run variation.
If your requirement is local inference
Liquid AI’s broader LFM catalogue is a separate path to investigate; it should not be confused with d1 availability. Verify the exact model and its current documentation before installing anything. In particular, check that its weights are actually downloadable, that the license permits your intended use, and that the documented runtime and hardware match your device.
Liquid AI’s Pipette documentation also cautions against reading a device-performance display as proof that a quality evaluation ran on that device. A quality score shown beside phone performance does not, by itself, mean the quality test was performed on the phone.
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.
