Use Liquid AI’s d1 when a task has a bounded answer—such as yes/no, one choice from a known set, or a score—and your application can act on probabilities instead of generated prose. Use a vision-language model (VLM) when the task needs flexible image interpretation, an explanation, or a natural-language response. d1 can process images, so the key distinction is not whether a model sees images; it is whether the job is a constrained decision or an open-ended answer.
What d1 returns—and what “zero output tokens” means
Liquid describes d1 as a decision model: it evaluates an input and returns probabilities over a fixed set of possible answers in one forward pass, rather than generating output tokens. The result is structured for an application to consume directly. Liquid’s documentation divides questions into three types:
As an Amazon Associate I earn from qualifying purchases.
- Noul: a yes/no decision with a probability from 0 to 1, such as “Is this message spam?”
- Choice: probabilities over named alternatives, such as which department should handle a support ticket.
- Score: a position on an ordered rubric, such as how urgent an issue is.
An application can ask multiple questions about the same input. That is useful when software needs to filter, route, rank, or select an action—not when the user needs a conversational explanation. See Liquid’s decision-model documentation and d1 launch article.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →How d1 differs from a vision-language model
Liquid’s model catalog separates decision models, built for classification, routing, and scoring across fixed outcomes, from VLMs, which work with vision and text inputs and outputs. A VLM or other generative model is usually the more natural fit when a task calls for a description, summary, explanation, or answer that cannot be specified in advance as a short list of outcomes.
#1 Best Overall
![Arduino® UNO™ Q 4GB [ABX00173]- Hybrid Board, Qualcomm Dragonwing QRB2210 microprocessor (MPU) & STM32U585 Microcontroller(MCU), AI Vision, Voice, IoT, Robotics, Linux Debian OS, Wi-Fi 5, USB-C](https://m.media-amazon.com/images/I/51cl4xygwGL._SL160_.jpg)
- Dual-Brain Hybrid Power: Combines the Qualcomm Dragonwing QRB2210 MPU (Quad-core Arm Cortex-A53 @ 2.0 GHz CPU, Adreno GPU, AI acceleration) and the real-time, low-power STM32U585 MCU for advanced applications like object recognition, voice commands, and motion detection.
- AI & Linux Capabilities: Unlocks AI-powered vision and sound solutions; runs Linux Debian OS for coding in Python and supports the Arduino ecosystem with libraries and Sketches; quick start with Arduino App Lab.
- Advanced Features: Equipped with 4 GB LPDDR4 RAM, 32 GB eMMC built-in storage, ideal for single-board computer (SBC) mode, running multiple simultaneous high-level processes, more complex AI or ML models, extensive logs. Dual-band Wi-Fi 5 (2.4/5 GHz), Bluetooth 5.1, and high-speed headers for vision, audio, and display peripherals.
- Seamless Expansion & Connectivity: Features the classic UNO form factor for shields compatibility, an 8x13 LED matrix, and a Qwiic connector for easy expansion with Modulino nodes; power and connect via the USB-C connector.
- Intended Use & Development: The perfect platform for prototyping robotics or IoT projects, empowering innovators with a unified development experience to mix Arduino Sketches, Python scripts, and containerized AI models in a single interface.
The categories are not mutually exclusive in their underlying technology. Liquid says d1-3B is based on LFM2.5-VL-3B, while d1-omni-600M is based on an encoder with vision and audio components. The practical choice is about the output contract: can the application use a probability distribution over known answers, or does it need generated language?
When bounded decisions are useful
d1 is worth testing when the possible outcomes are known and a structured answer can trigger the next step. Liquid’s launch article demonstrates several such applications:
- Classification and filtering: identifying cancellation intent or filtering support tickets. Liquid’s Smart Filter example compares its results with hand labels across 150 tickets; this is a specific demonstration, not a general accuracy guarantee.
- Routing and filing: choosing a support department or placing documents into folders and subfolders.
- Search support: locating relevant code in a repository or organizing search questions into a folder structure. These examples do not show that d1 replaces general-purpose code-search systems.
- Action selection: choosing the next available action in a demonstrated flight-search web agent.
- Visual inspection: sorting good and defective circuit boards, candles, cashews, and chewing gum in four VisA tasks. Liquid reports 85–97% accuracy and says d1 was not specifically trained for those inspection tasks. Those results do not establish performance on other factories, defect types, cameras, or datasets.
- Interactive screenshots: Liquid reports that adding a Tetris screen increased its score from 70 to 81 lines cleared, and that it solved 12 of 12 Wordle games in an average of 3.8 guesses using screenshots. These are company demonstrations, not independent benchmarks.
- Context selection: a coding-agent context-compaction demonstration removed 52% of tokens while retaining outputs needed for the next task, under Liquid’s stated sessions and setup.
These examples are starting points for evaluation, not evidence that d1 is ready to make an unattended, consequential decision. For a production gate, measure errors on representative data, decide what probability thresholds trigger action, define a fallback for uncertain cases, and retain human review when a false decision could cause material harm. Liquid’s cited materials do not establish a universal threshold or production accuracy guarantee.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsWhich d1 models and inputs are available?
Liquid’s October 7, 2026 release names two open-weight models. It says both are available on Hugging Face and have day-one llama.cpp support. Liquid’s October 5 launch article also describes a hosted d1 model through its API.
Rank #2
- Dual-Brain Hybrid Power: Combines the Qualcomm Dragonwing QRB2210 MPU (Quad-core Arm Cortex-A53 @ 2.0 GHz CPU, Adreno GPU, AI acceleration) and the real-time, low-power STM32U585 MCU for advanced applications like object recognition, voice commands, and motion detection.
- AI & Linux Capabilities: Unlocks AI-powered vision and sound solutions; runs Linux Debian OS for coding in Python and supports the Arduino ecosystem with libraries and Sketches; quick start with Arduino App Lab.
- Advanced Features: Equipped with 2 GB LPDDR4 RAM, 16 GB eMMC built-in storage, ideal to develop in PC-connected mode, running the OS, Python scripts, and basic network services (SSH) without a demanding GUI or heavy multitasking; great for lightweight AI and memory-optimized TinyML applications, needing local storage for basic OS and core libraries. Dual-band Wi-Fi 5 (2.4/5 GHz), Bluetooth 5.1, and high-speed headers for vision, audio, and display peripherals.
- Seamless Expansion & Connectivity: Features the classic UNO form factor for shields compatibility, an 8x13 LED matrix, and a Qwiic connector for easy expansion with Modulino nodes; power and connect via the USB-C connector.
- Intended Use & Development: The perfect platform for prototyping robotics or IoT projects, empowering innovators with a unified development experience to mix Arduino Sketches, Python scripts, and containerized AI models in a single interface.
| Model | Basis and stated inputs | Status noted by Liquid |
|---|---|---|
| d1-3B | Based on LFM2.5-VL-3B; accepts text and images. | Open-weight release; Liquid also describes a hosted API model. |
| d1-omni-600M | Based on LFM2.5-Encoder-350M; accepts text plus image or text plus audio. | Experimental checkpoint under active development. |
For service integrations, availability may differ by provider and modality. The October 5 article said Vercel and OpenRouter access was text-only at that time, with vision planned later. Check the current Liquid AI launch details and provider documentation before building around a particular endpoint.
What Liquid’s benchmark results do—and do not—show
Liquid reports a score of 48.57 on Decision Index v0.2.1’s public split for d1-3B in its October 7, 2026 release. Liquid says this is ahead of every model below 10 billion parameters and on par with Decider 35B-A3B on that benchmark. The release is a vendor report, not an independent evaluation.
On a separate table of seven public text benchmarks—SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI, and PAWS-X—Liquid reports a mean of 82.9 for d1-3B and 78.4 for d1-omni-600M. The same table reports 81.1 for Decider 4B and 77.1 for Decider 2B. A mean across different tasks can conceal weaknesses on an individual task, so it cannot establish how well a model will perform on your own workload.
Liquid also reports a comparison with GPT-6.1 Sol and Claude Opus 5.5: d1 matched or beat GPT-6.1 Sol on four of six applications, cost 19 to 200 times less, and answered faster on every task. Liquid says it ran each application once on October 5, 2026, using its d1 Playground comparison script. The chat models received one chat message and JSON output at default reasoning settings; costs used list prices without prompt-cache discounts, while d1 costs were calculated at $0.04 per million input tokens. The Smart Filter run covered 150 tickets and Smart Folders covered 105 passages; several code and compaction questions were written after d1’s pipeline was set. This is a one-run, vendor-reported snapshot with task-specific conditions, not a general price or quality guarantee. Details are in Liquid’s launch comparison.
Rank #3
- Single core ARM Cortex-A7 32-bit core, integrated with NEON and FPU
- Built in Micro's self-developed 4th generation NPU, with high computational accuracy and support for mixed quantization of int4, int8, and int16. Among them, int8 has a computing power of 0.5 TOPS and int4 has a computing power of up to 1.0 TOPS
- Built in self-developed 3rd generation ISP3.2, supports 4 million pixels, and supports various image enhancement and correction algorithms such as HDR, WDR, and multi-level denoisin
- It has powerful encoding performance, supports intelligent encoding, adapts to save bit rates according to the scene, and saves more than 50% of the bit rate compared to conventional CBR mode, making the captured images high-definition, smaller in size, and doubling the storage space
- The design with built-in RISC-V MCU supports low-power fast startup, 250ms fast capture, and simultaneous loading of AI model library, enabling facial recognition to be completed within 1 second
For vision, Liquid says d1-3B retained the vision capabilities of its LFM2.5-VL-3B backbone on standard vision benchmarks, but its October 7 release does not report the private vision split. Liquid also describes dedicated audio decision benchmarks as an open problem. The public text scores therefore should not be read as evidence of general vision or audio superiority. See the open-weight release.
How fast is d1 on edge hardware?
Liquid’s October 7, 2026 release reports the following d1-3B single-question latencies on selected devices:
| Device | Reported latency |
|---|---|
| NVIDIA Jetson Orin Nano | 50 ms |
| NVIDIA Jetson AGX Thor | 16 ms |
| NVIDIA Jetson AGX Orin 64 GB | 26 ms |
| Apple M5 Pro | 30 ms |
| NVIDIA RTX 4090 | 8 ms |
These are Liquid-reported measurements for one question, not a promise for every deployment. The same release reports 1,640 ms on Jetson Orin Nano for a 3.4K-token state and 202 ms for a 384-pixel image. State length, image resolution, number of questions, runtime, quantization, batching, and warm or cold execution can all affect an end-to-end comparison. Liquid also demonstrates d1-3B served on Jetson hardware in an Isaac Sim setup with NVIDIA collaboration. Its selected device results do not establish performance for every configuration; see the release methodology and deployment notes.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Hosted API or open weights?
The hosted API and open-weight routes have different deployment and billing implications. In its October 5 article, Liquid said API billing was based on input tokens with no output tokens. Under the pricing method stated there, images counted as 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image counted as 1,536 input tokens. These are published service details from that article, not a guarantee of current pricing; verify current terms before estimating cost.
Rank #4
- 【POWERFUL ESP32‑S3 CONTROLLER】Built‑in Xtensa 32‑bit LX7 dual‑core processor, 512KB SRAM, 8MB PSRAM, 16MB Flash for stable AI voice computing and multitask processing.
- 【Preloaded Dual AI Platforms】Comespre-installed with complete Deepseek and OpenAI voice dialogue projects.Experience intelligent voice interaction instantly. (Note: OpenAI functionality requires your own API key.)
- 【STABLE WIRELESS & CLEAR AUDIO】Integrated 2.4GHz Wi‑Fi + Bluetooth 5 (LE); dedicated audio decoding module for natural, responsive voice interaction.
- 【USER‑FRIENDLY VISUAL & PLUG‑AND‑PLAY】2” TFT‑SPI color screen shows real‑time chat; modular design, no extra wiring, ready to use after setup.
- 【FULL LEARNING SUPPORT】45 programmable GPIOs, rich interfaces, online web tutorials, free technical support for beginners & developers.
Open weights may allow local deployment, but the application team still has to validate its hardware, runtime, security, privacy, and operational requirements. Liquid promotes on-device use; that alone does not establish that a particular deployment meets your privacy or compliance needs.
A practical way to choose
Compare candidates on the actual task rather than the “zero tokens” label. Use labeled examples from the intended workload and the hardware and service path you plan to deploy.
| Decision axis | Question to answer |
|---|---|
| Output shape | Are valid answers yes/no, one of a known set, or an ordered score—or does the task need arbitrary text? |
| Input modality | Does the workload use text, images, or audio, and does the chosen model and deployment support that exact combination? |
| Task quality | On representative labeled examples, what errors occur? Are probabilities calibrated well enough for the thresholds you need? |
| Latency | What is end-to-end latency with the actual state length, image size, batch size, runtime, and target device? |
| Integration | Can the application consume probabilities, or does it need explanations, tool use, or conversational turns? |
| Cost and privacy | What are the current API charges or hardware costs, and where does the data travel in the deployment you will actually use? |
A two-stage design is also worth testing: let d1 handle high-volume, bounded decisions, then send uncertain or open-ended cases to a VLM. This is an architectural option suggested by the models’ output shapes, not a performance result established by Liquid’s published comparisons.
Recommended Free Tools
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.
