Liquid AI Releases d1-3B and d1-omni-600M Decision Models for Edge Inference

Liquid AI published two downloadable decision models on October 7

Liquid AI announced d1-3B and d1-omni-600M on October 7, 2026. Both repositories are public and ungated on Hugging Face. The release targets structured decisions at the edge: an application supplies a state plus named questions, and the model returns typed yes/no probabilities, a choice among declared options or a score. It does not produce an open-ended text answer.

That makes the launch narrower than a general multimodal assistant and potentially more useful for specific pipelines such as routing, moderation, reranking, visual inspection and agent guardrails. There is no hosted API price attached to the open-weight release. Teams supply their own hardware and absorb download, serving, monitoring and review costs.

d1-3B handles text and images in a single decision pass

The d1-3B repository identifies 3,123,483,888 BF16 parameters and a 32,768-token context window. It is post-trained from LFM2.5-VL-3B and accepts text, JSON, images or a mixture. Multiple named questions can share one encoded state, and the documented response includes typed answers, confidence or probability fields and zero output tokens.

Zero output tokens does not mean zero compute or zero cost. The model still processes the input and decision schema, and local cost depends on hardware, utilization, batching and operational overhead. It also does not mean a decision is correct or calibrated for a new domain. Teams need workload-specific thresholds, abstention behavior and consequence-weighted evaluation.

d1-omni-600M adds audio but remains an experimental release

d1-omni-600M contains 587,161,089 FP32 parameters: Liquid AI describes a shared trunk and decision head plus separate vision and audio encoders. It supports text with images or text with one 16 kHz mono audio clip, but not images and audio in the same request. Audio is capped at 30 seconds and was trained around English-speaker requests to an assistant.

Its model card documents a 16,384-position combined limit. With images, state and question text is cut to 896 tokens as trained. Liquid AI publishes no latency numbers for this model because it is an early research release under active development. The card also warns that bfloat16 changed the top answer on some of its checks and recommends float16 on GPUs after its own comparison against float32.

The benchmark claims are vendor-run and one table is inconsistent

Liquid AI reports a 48.57 score for d1-3B on Decision Index 0.2.1, calling it the best result below 10 billion parameters and above the listed Decider 35B-A3B score of 47.11. The model card says Liquid AI used the official scorer but did not submit the result to the public leaderboard. That is provider-reported evaluation, not an independent ranking.

The October 7 announcement reports an 82.9 mean for d1-3B across seven public datasets. The live d1-3B model card currently reports 77.1 across an eight-dataset table and differs on individual overlapping values, including SQuAD 2.0, Civil Comments and PubMedQA. The d1-omni-600M card reproduces yet another view of the comparison. These may reflect updated evaluations, but the pages do not provide a dated changelog or reconcile the differences. Use pinned datasets, prompts, scorer versions and repository commits before making a quality claim.

The published speed numbers are warm, provider-measured results

For one question, Liquid AI reports 8 ms on an NVIDIA RTX 4090, 9 ms on an AMD MI325X, 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin 64 GB, 50 ms on a Jetson Orin Nano and 30 ms on an Apple M5 Pro. It also reports separate times for three questions, a 3,400-token state, a 384-pixel image and a packed batch of 64 states.

Those figures describe warm calls under Liquid AI’s setup. The card says GPU measurements used BF16 and medians of 20 runs; the RTX 4090 single-question row used compiled CUDA graphs, while an uncompiled call took 16 ms. First calls with new shapes pay compilation or kernel-selection overhead. There are no tail-latency, cold-start, power, memory, accuracy-at-latency or end-to-end application measurements.

Open weights do not mean unrestricted commercial rights

Both repositories use LFM Open License v1.0 rather than an OSI-approved software license listed by Hugging Face. The text grants broad copyright and patent permissions subject to conditions, including attribution and redistribution obligations. Its commercial-use clause says use by a legal entity with annual revenue of $10 million or more is not licensed under the agreement.

That threshold applies to the legal entity as defined by the license, including entities under common control. Qualified nonprofit organizations receive a stated exception only for non-commercial or research purposes. A public download therefore establishes artifact access, not production permission for every company. Teams near or above the threshold should obtain legal review and a separate Liquid AI agreement before commercial use.

Loading the models executes repository-supplied code

The documented examples require Transformers 5.14 or newer for d1-3B and 5.15 or newer for d1-omni-600M, then load each repository with trust_remote_code=True. That flag allows custom Python from the model repository to execute. Pin the repository commit, inspect the referenced Python and dependency graph, isolate the build and runtime, block unnecessary network and credential access, and retain hashes for every deployed artifact.

Do not place an unreviewed model or its custom code directly inside a privileged agent path. The model’s answer should be a proposal consumed by deterministic authorization, validation and logging. High-impact denials, financial decisions, safety controls and irreversible actions need independent checks and an explicit fallback or human-review route.

What builders should do next

Trial d1-3B when a workload already has closed answer choices and text or image inputs, and trial d1-omni-600M only when its experimental status, 30-second English-audio boundary and precision constraints fit. Freeze the exact commit, license record, code review and hardware image. Compare against a deterministic baseline and one established model on representative, shifted and adversarial cases.

Measure class-level precision and recall, calibration, abstention, severe-error rate, latency percentiles, memory, power, throughput and total serving cost. Evaluate every modality separately and test malformed schemas and missing inputs. Constrain use behind host-enforced policy; wait when rights or benchmark provenance are unresolved; reject deployment when the model decision itself would become permission for a consequential action.

Disclosures

AccessAllGPT received no Liquid AI account, API access, hardware, weights, code, credits, benchmark data, briefing, legal guidance, review, payment or compensation for this article. Liquid AI, Hugging Face, NVIDIA and AMD did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank the models or hardware. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Liquid AI, Hugging Face or organizations cited. Publication-wide relationships are listed on the disclosures page.