Choose Liquid AI’s d1 when an edge application needs a bounded decision—such as a label, yes/no result, or score. Choose a generative small language model (SLM), such as Liquid’s LFM2.5-1.2B-Instruct, when it needs to produce flexible text, explanations, summaries, or instruction-following responses. The key difference is the output: d1 returns probabilities for specified answers; an SLM generates tokens.
How d1 differs from a generative SLM
Liquid AI describes d1 as a decision model, not a general-purpose text generator. Given a state—text, an image, or both, depending on the model—and one or more structured questions, it returns probabilities for possible decisions in a single forward pass. Its October 5, 2026 announcement names yes/no questions (“noul”), choosing a label from options, and scoring on a scale. Liquid’s October 7 open-weight announcement explicitly contrasts d1 with its generative models: d1 does not produce tokens.
As an Amazon Associate I earn from qualifying purchases.
A generative SLM instead produces text. That makes it a more natural fit when the application cannot fully specify its output in advance, or when a person needs a written answer. The distinction is about the shape of the task, not a claim that either family is universally more capable.
Recommended Free Tools
| Decision point | d1 decision model | Generative SLM |
|---|---|---|
| Output | Probabilities over declared decision answers | Generated text, including instruction-following responses |
| Natural task shape | Classification, yes/no checks, scoring, filtering, routing, inspection, bounded action selection | Chat, summaries, explanations, RAG answers, flexible extraction, tool-use responses |
| Useful evaluation | Decision Index or task-specific accuracy, precision/recall, or F1 against labeled examples | Task- and use-case-specific quality tests, plus relevant generation benchmarks |
| Performance measure | Time per decision or state, including context and image size | Prefill and decode rates, plus full response time and output length |
| Deployment considerations | Model footprint, input size, runtime, and batching or packed-state behavior | Parameter count, quantization, context length, memory, runtime, and modality |
Which model fits your edge application?
Use d1 for decisions with a defined answer space
Consider d1 when the application can ask a question such as “Is this item damaged?”, “Which of these routes applies?”, or “How urgent is this alert on a specified scale?” A bounded answer makes it possible to inspect the model’s probabilities and evaluate the decision against labeled cases. Potential uses include filtering inputs before a larger system handles them, classifying images, routing requests, and selecting among predefined actions.
#1 Best Overall
- EVOLUTION AMD RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Plan the interface around explicit questions and answer choices. If users need a rationale in natural language, d1’s decision output alone does not supply one; that requirement may call for a separate generation step or a generative model. Do not treat a probability as a guarantee of correctness: calibrate thresholds and test error costs for the application.
Use a generative SLM when the output must be language
Choose an SLM when the device needs to draft a response, summarize a passage, explain a result, answer a question in natural language, or follow flexible instructions. Liquid’s LFM2.5-1.2B-Instruct is one example, but the exact model and version matter: Liquid’s LFM2 documentation lists 350M, 700M, 1.2B, and 2.6B sizes, while the LFM2.5 release includes Base, Instruct, Japanese, vision-language, and audio-language models.
A generative model can also be prompted to return structured output, but its output is still generated text that an application may need to parse and validate. If the task is simply to choose among a fixed set of labels, a decision-model interface is more directly aligned with that requirement.
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 minuteUse both only when the product needs both jobs
A pipeline can use a decision model to filter or route an input and a generative model to produce a user-facing explanation or response. That adds integration, latency, and failure modes, so it is worth doing only when each stage solves a distinct requirement. Measure the complete pipeline rather than assuming that combining models will be faster or more accurate.
Rank #2
- Built for Local AI Development: AMD Ryzen AI Halo is designed for local AI development and inference, featuring 128GB unified memory and support for up to 200B parameter models to build and run intensive AI workloads locally.
- 128GB Unified Memory: Features 128GB LPDDR5x unified memory at 8000 MT/s with 256 GB/s memory bandwidth, providing a shared memory pool across the CPU, GPU, and NPU to support larger AI models.
- AMD Ryzen AI Max+ 395 Processor: Features 16 cores, 32 threads, and Zen 5 architecture, paired with AMD Radeon 8060S integrated graphics featuring 40 RDNA 3.5 compute units and an AMD XDNA 2 NPU with up to 50 TOPS.
- Linux AI Developer Platform: Purpose-built for Linux-based AI development with full AMD ROCm software support and preloaded tools, models, and workflows optimized for local AI development.
- Compact, Connected Design: Includes a 2TB M.2 SSD, 10GbE LAN, Wi-Fi 7, Bluetooth 5.4, USB-C connectivity, and HDMI 2.1b.
Can d1 run on an edge device?
As of Liquid AI’s October 7, 2026 announcement, open-weight d1-3B and experimental d1-omni-600M were available on Hugging Face, with day-one llama.cpp support. Liquid describes d1-3B as accepting text and images. The company describes d1-omni-600M as handling text plus either images or audio and labels it an early research release under active development. Confirm the current model files, runtime support, and modality details before choosing a deployment.
Liquid’s October 5 announcement described d1 API access and availability through Vercel and OpenRouter for text at that time. That is a different deployment route from running open weights locally, and the announcement date and modality matter: availability should not be assumed to be uniform across models or input types.
Liquid-reported d1-3B latency
The figures below are Liquid AI’s reported measurements, published October 7, 2026. They are starting points for evaluation, not guarantees for other hardware configurations, runtimes, or inputs. The 3.4K-token state and 384px image columns measure different input conditions from the one- and three-question columns.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute| Platform | One question | Three questions | 3.4K-token state | 384px image | 64 packed states |
|---|---|---|---|---|---|
| Apple M5 Pro | 30 ms | 41 ms | 640 ms | 62 ms | 78/s |
| NVIDIA Jetson AGX Thor | 16 ms | 20 ms | 220 ms | 35 ms | 262/s |
| NVIDIA Jetson AGX Orin 64 GB | 26 ms | 35 ms | 560 ms | 83 ms | 110/s |
| NVIDIA Jetson Orin Nano | 50 ms | 73 ms | 1,640 ms | 202 ms | 38/s |
These results illustrate why “edge latency” needs an input definition: a longer state or an image can take substantially longer than a single short question. For a real product, measure the full path from input preparation through model output, at the device’s expected load and with representative states.
Rank #3
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
What the published benchmarks can—and cannot—tell you
Liquid AI reports d1-3B scoring 48.57 and d1-omni-600M scoring 15.95 on the Decision Index v0.2.1 public split in its October 7, 2026 release. Liquid says d1-3B is ahead of every model under 10B and on par with Decider 35B-A3B on that index. These are company-reported decision-model results; they do not establish that d1 outperforms generative SLMs at text generation.
Do not compare a Decision Index score directly with an SLM score on MMLU, IFEval, or GSM8K. Those benchmarks cover different task families. Keep model generation and benchmark versions separate as well: Liquid AI’s 2025 LFM2 technical report reports 79.56% on IFEval and 82.41% on GSM8K for LFM2-2.6B. Those figures are for that reported LFM2 model, not d1 or LFM2.5.
Liquid’s October 5, 2026 d1 announcement also describes a six-application comparison with GPT-6.1 Sol and Claude Opus 5.5. The company says d1 matched or beat GPT-6.1 Sol on four tasks and was 19x to 200x cheaper than both models. The stated method ran each application once on October 5, used default reasoning settings and list prices without cache discounts, and applied task-specific scoring. Treat these as results for the selected applications and stated method, not a general quality or cost guarantee.
Free tools Windows power users keep installed
One-click scans. No signup required.
The published material reviewed here does not establish an independent, identical-task, same-hardware comparison of d1 against a broad set of generative SLMs. Choose based on the required output, then compare candidates on your own workload rather than using scores from incompatible benchmarks as a proxy.
How to evaluate candidates on your hardware
- Specify the output contract. Write down whether each request needs a fixed decision, a confidence or score, or generated language. Define the labels, allowed choices, and what happens when confidence is low.
- Build a representative test set. Include normal cases, edge cases, difficult inputs, and the text or images the device will actually process. For decisions, use trusted labels; for generation, define what counts as a useful and safe response.
- Run each candidate with its intended runtime and settings. Record model version, quantization, device, runtime, input length, image size, and batch or packed-state configuration. A result without those details is difficult to reproduce or apply elsewhere.
- Measure the metric that matches the task. For decision models, track task quality and end-to-end time per state. For generative SLMs, measure output quality, prefill and decode behavior, full response time, and memory use. Include input processing and any application-side parsing.
- Test deployment constraints. Check memory headroom, sustained performance, supported modalities, offline behavior, and the complete data flow. Local execution can avoid sending state to a cloud service when the application is configured locally, but verify that other components do not transmit it.
- Set an acceptance threshold before choosing. Compare quality, latency, memory, and failure behavior against the product’s requirements. Keep a fallback or abstention path for uncertain or unsupported inputs where errors have meaningful consequences.
Deployment options and version cautions
Liquid’s LFM2 documentation describes CPU, GPU, and NPU hardware support. The LFM2.5 release announcement names llama.cpp, MLX, vLLM, and ONNX, and says it supports CPU and GPU acceleration across Apple, AMD, Qualcomm, and Nvidia hardware. These family-level statements do not guarantee identical support for every model, device, or runtime combination; check the current documentation for the specific target.
Liquid’s LEAP platform page currently begins with a deprecation notice. Do not assume LEAP is required or is the default path for deployment. Likewise, distinguish LFM2 from LFM2.5 and identify the particular model variant when interpreting benchmark or throughput claims.
A device-specific generative throughput example
On a Samsung Galaxy S25 Ultra CPU using llama.cpp Q4_0, Liquid AI reports 70 decode tokens per second for LFM2.5-1.2B-Instruct and 40 decode tokens per second for Qwen3-1.7B, with reported memory use of 719 MB and 1,306 MB respectively. This is a vendor-run result for that setup; it should not be extrapolated to other devices, quantizations, or workloads.
Make the choice by task, then verify it locally
For a fixed choice, score, filter, or route, start with d1 and validate its decisions against representative labeled inputs. For open-ended language, start with a generative SLM and test response quality as well as runtime behavior. If the product needs both, benchmark the complete multi-model path. The available figures are useful for identifying candidates, but the final choice depends on task-specific quality and performance on the hardware and inputs you will actually ship.
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.
