Liquid AI released two open-weight decision models on October 7, 2026: d1-3B, which accepts text and images, and the experimental d1-omni-600M, which adds an audio input path. The company positions them for classification, routing and scoring on local hardware rather than free-form conversation. 1 2
Downloadable weights and upstream inference support give developers concrete ways to try the release. The llama.cpp project merged d1-3B vision-decision support on October 7. Those artifacts establish availability, however—not independent confirmation of Liquid’s performance claims. 3 4
Decisions, not conversational answers
Liquid documents an interface that returns probabilities or selections for three question types: binary questions, named choices and ordered ratings. Its intended uses include ticket routing, moderation, visual inspection and spoken-intent classification. The company directs free-form generation and complex multistep reasoning to language models instead. 2
According to Liquid’s model cards and documentation:
- d1-3B has 3.12 billion parameters, accepts text and images, and has a nominal 32,768-token context. It is Liquid’s recommended model of the two. 5 2
- d1-omni-600M has 587 million parameters and a nominal context of 16,384 combined positions. It accepts text with images or text with audio, but cannot combine image and audio inputs. Liquid labels it an experimental research release. 6
The smaller model has important modality-specific limits. With images, its state and question text is truncated to 896 tokens. Its audio training covers English speaker-to-assistant requests, with training clips capped at 30 seconds. The card describes temperature calibration for text answers, while image and audio answers use trained softmax outputs; equivalent confidence calibration across modalities is not established. 6
The downloadable models are also separate from Liquid’s hosted d1 service. The API’s application results, pricing and latency should not be carried over to these checkpoints. 2
What the edge latency figures show
Liquid reports the following d1-3B timings. Each column represents a different workload, so the figures are not interchangeable measures of multimodal response time. 1
Hardware One question 384-pixel image 3.4K-token state Jetson AGX Thor 16 ms 35 ms 220 ms Jetson AGX Orin 64 GB 26 ms 83 ms 560 ms Jetson Orin Nano 50 ms 202 ms 1,640 ms NVIDIA RTX 4090 8 ms 17 ms 102 msIn particular, Thor’s 16 ms result is not its image-processing result. Liquid’s model card specifies warmed-up calls with one request at a time. The RTX 4090’s 8 ms result uses compilation and CUDA graphs; Liquid reports 16 ms without that optimization. 1 5
No inference figures are published for the experimental omni checkpoint in the reviewed material. Its lower parameter count therefore does not establish that it is faster or more energy-efficient than d1-3B. 6
Outside testing found in the research is narrower: a Reddit practitioner reports 6 ms median latency for a self-quantized d1-3B on an RTX 4090 in a single-word classification experiment. That is not a replication of Liquid’s Jetson multimodal workloads. 7
Quality comparisons remain unsettled
Liquid reports Decision Index 0.2.1 scores of 48.57 for d1-3B and 15.95 for d1-omni-600M, compared with 40.70 for Decider 4B and 28.97 for Decider 2B. The company says it used the official scorer, but the runs were not leaderboard submissions. Decision Index is a chance-corrected composite, not a simple accuracy percentage. 6 8
The choice of benchmark changes the comparison. Liquid’s selected seven-text-task mean puts omni at 78.4 versus Decider 2B’s 77.1, even though omni trails substantially on Decision Index. By the research cutoff of October 12, the retrieved benchmark repository identified revision 0.3.2, newer than the edition used in Liquid’s release comparisons. 6 9
Liquid’s claim that d1-3B leads decision models below 10 billion parameters is not established by those results. StartLux’s own results report higher scores on Decision Index 0.2.1: 52.75 for its 4B model and 58.63 for its 9B model. StartLux also describes its runs as unsubmitted. These are competing self-reported evaluations, not an independently settled ranking. 1 10
There are also unresolved discrepancies within Liquid’s publications. Its employee-authored Hugging Face announcement gives d1-3B scores of 83.3 on SQuAD 2.0 and 68.3 on PubMedQA; its corporate announcement gives 85.3 and 66.0. Both report a seven-task mean of 82.9. The retrieved pages do not explain the mismatch. The model card’s eight-task mean of 77.1 includes HelpSteer2 and is a different aggregate, not a directly comparable replacement. 11 1 5
Visual evaluation does exist, despite the Hugging Face announcement saying it reports no vision or audio benchmarks. The retrieved d1-3B card lists 11 public image benchmarks, with a Liquid-reported mean of 74.1 versus 73.9 for the backbone and at most 1,000 examples per benchmark. Those results provide vendor-generated visual evidence, not independent multimodal validation. 11 5
Open weights, with commercial limits
Both checkpoints identify the LFM Open License v1.0. The license imposes redistribution obligations and does not grant commercial-use rights to businesses above its approximately $10 million annual-revenue threshold. That qualifies the corporate announcement’s unrestricted-deployment language: open-weight is the appropriate description, not unrestricted open source. Explicit commercial terms for businesses above the threshold remain unresolved in the reviewed material. 12 6 1
What developers still need to establish
The practical opportunity is narrower than replacing a general-purpose assistant: a fixed-choice interface could remove generated-answer parsing from some classification pipelines. Liquid’s guide also demonstrates falling back to a stronger model when confidence is low. Whether that helps a particular application depends on validating decision quality, confidence thresholds and failure handling on its actual workload. 2 13
Important follow-up questions remain: which checkpoint hashes, evaluation revisions and raw outputs explain the conflicting text scores; whether Liquid will submit both models to the current Decision Index; and what cold-start latency, tail latency, peak memory, power consumption, thermal conditions and precision settings look like on edge hardware. The reviewed timing results do not resolve those deployment questions. 1 11 5 9 6
Independent tests are also needed for noisy speech, non-English audio, difficult images and out-of-distribution inputs, particularly for experimental omni. Until then, the release provides an implementation path for structured local decisions, but its broader reliability and efficiency remain workload-specific questions. 6 13
Research basis: a supplied source review with an October 12, 2026 cutoff. The underlying research did not execute models or reproduce benchmarks.