Liquid AI Releases Open D1 Decision Models for Edge Devices

Executive Summary

• Liquid AI released two open-weight decision models in its d1 family, d1-3B and d1-omni-600M, both available on Hugging Face as of October 7, 2026.
• d1-3B scores 48.57 on the Decision Index 0.2.1, which the team describes as the best result for any decision model under 10B parameters, ahead of Decider 35B-A3B at 47.11.
• The smaller d1-omni-600M is an experimental research release that handles text with either images or audio, while d1-3B handles text and images, per the release post.
• Speed results produced with NVIDIA show d1-3B answering a single question in 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin, and 50 ms on a Jetson Orin Nano.

Key Takeaways

• Liquid AI is positioning the d1 family against larger decision models rather than general-purpose language models, using its Decision Index and seven public benchmarks as the comparison base.
• d1-3B and d1-omni-600M answer in a single forward pass instead of generating tokens, which is the mechanism behind the reported millisecond latencies.
• d1-omni-600M adds audio as an input modality but remains an early research release, and Liquid AI did not publish speed figures or vision and audio benchmark scores for it.
• Deployment code requires transformers 5.14 or later plus trust_remote_code, a practical constraint for teams with pinned dependency stacks.

Hugging Face Report Details the Two-Backbone Decision Model Architecture

The d1 models are built on Liquid AI's Liquid Foundation Models but diverge from generative models in output behavior. Rather than producing tokens, they answer in a single forward pass, which the release post presents as the core design choice behind their latency profile. The two models are not variants of one backbone. d1-3B trains from LFM2.5-VL-3B, a decoder-only vision-language model that takes text and images. d1-omni-600M trains from LFM2.5-Encoder-350M, a bidirectional encoder, and adds vision and audio encoders to accept either text and image or text and audio. Liquid AI describes d1-omni-600M as an early research release that is still under development, a caveat that separates it from the more complete d1-3B.

Hugging Face Report Shows d1-3B Leading Sub-10B Decision Index

On the Decision Index 0.2.1, d1-3B scores 48.57, which the team says is ahead of every 4B and 9B model and of Decider 35B-A3B at 47.11. The claim matters most for buyers weighing parameter count against decision quality, since a 3B model is being placed above a 35B model on that index. Liquid AI also benchmarked both models on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding. d1-3B posts a mean score of 82.9, the highest in the table and above Decider 4B at 81.1. d1-omni-600M scores 78.4, ahead of Decider 2B at 77.1 with roughly a quarter of the parameters.

Hugging Face Report Documents Multimodal Benchmark Splits

The per-dataset results are not uniformly favorable. On SQuAD 2.0, d1-3B reaches 83.3 against Decider 4B at 76.0, and on PAWS-X it posts 79.5 against 69.8. On BoolQ and XNLI, however, Decider 4B leads, 89.0 against 86.3 and 88.6 against 85.6 respectively. d1-omni-600M shows a similar pattern, leading on Civil Comments at 95.8 and PAWS-X at 79.5 while trailing Decider 2B on BoolQ at 77.7 against 87.3. Benchmark composition therefore matters when interpreting the mean scores. Liquid AI states it validated that d1-3B retains the vision capabilities of its backbone and that d1-omni-600M handles all three modalities, but it reports no vision or audio benchmarks, citing a private vision split in Decision Index v0.3 and describing audio decision benchmarks as an open problem.

Hugging Face Report Details Edge and GPU Latency Across NVIDIA and AMD Hardware

The speed evaluation was run with NVIDIA across the GeForce RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano, with AMD MI325X results also included in the GPU table. Liquid AI reports that d1-3B answers a single question in under 50 ms on every measured device, and that three questions take only 1.3 times the time of one, with the AGX Thor moving from 16 ms to 20 ms. On the GPU side, d1-3B answers a question in under 10 ms and processes a 384px image in under 18 ms on both platforms. Throughput on packed batched states differs sharply by device, from 38 states per second on the Jetson Orin Nano to 262 on the AGX Thor and 1,106 on the AMD MI325X. No speed numbers were published for d1-omni-600M.

Related: How Sustainability Drives ROI in 2026, According to McKinsey and BCG

Hugging Face Report Outlines Deployment Path and Coding Pattern

Both models are open-weight and available on Hugging Face, with installation requiring transformers 5.14 or later alongside torch, torchvision, and pillow. Liquid AI notes the models ship their own code and must be loaded with trust_remote_code set to true. The published example shows named questions defined as typed tasks, including boolean checks, choice selection with named criteria, and scored urgency levels, then answered together over one text state in a single pass. A second example passes an image as the whole state, and a third batches multiple requests with no padding. A System One Arcade demo space is also available for testing.

For deeper context, see our Robotics analysis: "How Robotics Is Shifting From Pilots to Core Infrastructure".

Hugging Face Implementation Risks

The source does not document model licensing terms, security review, hardware requirements beyond those tested, or certification for regulated workloads, so buyers cannot assess compliance fit from this post alone. The benchmark evidence rests largely on Liquid AI's own Decision Index and a seven-dataset table, and the source says no vision or audio benchmarks are reported, leaving the multimodal claims unquantified. d1-omni-600M carries a research-stage label with no latency figures, which limits production planning. Loading with trust_remote_code executes repository-supplied code, a supply-chain consideration that teams should weigh against their dependency and review policies. The 16 ms to 50 ms figures are device-specific measurements, not guarantees across other hardware or longer contexts.

Additional coverage: Portal Space Systems, Geodesic & ARK Invest Target Orbital Propulsion in...

What This Means for Practitioners

Teams evaluating fast classification, routing, or triage at the edge now have an open-weight option that trades token generation for a single forward pass, which changes the cost calculus for high-volume inference on constrained devices. The practical decision is less about headline scores than about task match. d1-3B leads on several datasets but trails Decider 4B on BoolQ and XNLI, so buyers with those workloads should benchmark on their own data before committing. Two constraints deserve early attention: the transformers 5.14 floor and trust_remote_code requirement may conflict with pinned or security-restricted environments, and d1-omni-600M should be treated as experimental until latency and audio benchmarks are published.