d1-Omni-2B: What Liquid AI's Quietest 2B Decision Model Ships With — and What It Leaves Out

The Hugging Face repository for LiquidAI/d1-Omni-2B was created at 19:30 UTC on 24 October 2026. The first Liquid AI blog post to mention the model by name went up on 27 October. For those three days the repo was public and complete — a full spec sheet, 4.33GB of BF16 safetensors, working Transformers code — and nothing the vendor published said it existed. When the mention finally arrived it was two paragraphs and a four-row table inside a six-model family launch, not a launch of its own. That is the shape of this release, and it is the honest frame for everything below: the weights are inspectable, the claims are the vendor's, and the parts a buyer would normally get from an announcement — a price, a hosted endpoint, an independent benchmark — are not there yet.

The five-day gap between the weights and the words

Liquid AI's decision-model story has been told twice, and d1-Omni-2B is in neither telling. On 7 October the company published "Open d1: Edge decision models for text, vision, and audio," which introduced exactly two open checkpoints: d1-3B, a 3.12B vision-and-text model, and d1-omni-600M, a 587M model that took text with images or text with up to 30 seconds of audio. That post is still live, and it still contains no 2B model.

The 2B line arrived on a different schedule. On 21 October the base checkpoint LFM3-Omni-2B appeared on Hugging Face. On 24 October, twenty-four seconds apart, Liquid AI created repositories for d1-2B and d1-Omni-2B. The LFM3 launch post on 27 October then named the family: six models around 2B parameters, four covering text, vision, audio and omni, and, in the post's own words, "the last two are decision models specialized from the vision and omni models." That sentence is the announcement. There has been no separate d1-Omni-2B post, no pricing page, no launch partner quote, and no update to the 7 October article.

So the model is not undocumented — the repository is one of the better-documented small checkpoints Liquid AI has published, and the vendor's docs page for decision models lists it alongside d1-2B and d1-omni-600M. It is under-announced. That distinction matters because it tells you which claims have been through a launch review and which have not.

Reading the card as a spec sheet

Strip the framing and the checkpoint is straightforward. d1-Omni-2B is a post-trained decision model built on LFM3-Omni-2B, and it is not a chat model: it writes no text and produces no speech, and every call returns output_tokens: 0.

• Parameters — 2.17B, shipped as a single 4.33GB BF16 safetensors file.

• Architecture — the LFM3 backbone: 24 layers, of which 13 are gated short convolutions, 5 are sliding-window attention with a 256-token window, and 6 are global attention. The family post describes this tri-hybrid mix as the result of a hardware-in-the-loop architecture search run across 14 device and backend combinations.

• Context and vocabulary — 128K tokens and a 128,000-token vocabulary shared with the rest of LFM3.

• Input — a state that can be text or any JSON value, plus pictures, video frames and audio clips, plus any number of named questions, written or spoken. The card's API takes all of those in one media list, which is a meaningful difference from the 587M checkpoint: d1-omni-600M is documented as text-with-images or text-with-audio, not both at once.

• Output — typed answers with probabilities. Yes/no, a pick from named options, an ordered score, a value read from a document with its span, a number with a p10–p90 interval, a bounding box, or every instance of an object with a count.

• Languages — sixteen. Speech covers the same set except Hindi, Thai, Vietnamese and Indonesian, which is worth flagging because the text side lists them.

• License — LFM Open License v1.0. Commercial use is granted below $10,000,000 of annual revenue; above that, the license text says commercial use "is not licensed under this Agreement."

Audio is the claim worth testing

The reason to reach for this checkpoint over d1-2B is audio. Liquid AI positions it for "call and voice-message routing and sound events," and the model card carries the one piece of evidence that a launch post would normally have led with: a spoken-versus-written comparison on the same items.

• Business routing and triage (DecisionBench, Hanno Labs v1) — 83.4 written, 78.2 spoken.

• Customer service cases (Typed Decisions) — 50.5 written, 48.0 spoken.

• JevBench v1.4.1, the rows a voice can say — 87.6 written, 89.0 spoken, meaning the spoken run scored slightly higher, which is noise rather than a gain.

• Sound events (ESC-50) — 84.5, with no written equivalent.

• Spoken language identification (FLEURS, 43 languages) — 92.5.

Those numbers support the pitch. The rest of the audio section is the part that tells you where not to build, and it is unusually candid for a vendor card. Spoken short states fall from 42.0 written to 22.8 heard. Belebele passages in eight languages fall from 83.6 to 41.8. The card states plainly that long spoken passages lose much of their accuracy, "most of all outside Latin-script languages," and directs anyone without audio input to d1-2B instead, which it says is stronger on text and single boxes, reports missing values more often and is better calibrated.

Read together, those rows describe a model that is good at short, routable utterances in major languages and much weaker at anything that requires holding a long spoken state. Call-centre triage fits. A forty-minute recorded meeting does not.

The scoreboard is the vendor's, and the index changed version

Every performance number on this model is Liquid AI's own. The card is explicit about method: all four d1 models were scored on the same items with the same scoring, so some figures differ slightly from the individual cards they came from. It does not say a third party has reproduced any of it, and no independent evaluation of d1-Omni-2B exists as of writing.

• Decision Index public suite, v0.3 — 49.0 for d1-Omni-2B, against 51.1 for d1-2B, 49.0 for d1-3B and 18.3 for d1-omni-600M.

• DecisionBench (Hanno Labs) v1 — 76.4, against 77.1, 73.1 and 58.7 respectively.

• decisionbench.ai v4 — 84.7, against 85.2, 83.6 and 57.9.

• ImajevBench v2.0-lite — 72.3, against 75.5, 64.0 and 21.7.

• The vendor's own board of more than a hundred benchmarks — 61.8, against 63.2, 58.8 and no score for the 600M.

• Vision average across 18 image benchmarks — 67.7, against 70.2, 65.4 and 42.8.

One trap sits in the middle of this. The 7 October post reported d1-omni-600M at 15.95 on Decision Index v0.2.1. The new card reports it at 18.3 on v0.3. Those are different suites and the difference is not a capability gain; do not subtract one from the other. The only fair comparison is the one the card actually made, with all four models re-scored on the same items.

The pattern that survives the re-scoring is that d1-Omni-2B sits just below d1-2B on nearly everything and above the previous generation by a wide margin. Liquid AI frames it as beating d1-3B on its decision board and on 16 of 18 image benchmarks at 70% of the parameters, which is true of the table as printed and is also a comparison against a model the family has now replaced.

What the repository does not answer

Five questions are still open, and none of them is answered by reading the weights.

• Hosted access and price. The LFM3 post says the base and post-trained models are available on Hugging Face, the Playground and the Liquid AI API. The decision-models docs table, however, lists d1-Omni-2B as open weights and reserves its hosted-only rows for d1.1 and d1-ultra, and the Hugging Face model page states that "this model isn't deployed by any Inference Provider." No per-call price for this checkpoint is published anywhere this piece could find, so treat hosted access as plausible but unconfirmed.

• Training data. The card gives no corpus, no composition, no dataset description. It mentions testing on clips and voices "that no training row has," which implies a training set exists and says nothing about what is in it.

• Quantisation. The main repository ships BF16, and Liquid AI publishes three sibling builds: MLX fp16, MLX 8-bit and an ONNX export. There is no vendor GGUF, though Ollama's d1-omni library carries a 2b tag at 4.4GB and 125K context, and the size is close enough to the BF16 checkpoint that it does not look like aggressive quantisation.

• Independent evaluation. None. Every number above is vendor-run with the vendor's own scorer on the vendor's own board.

• Whether it replaces d1-omni-600M. The card recommends d1-2B over this model when there is no audio, and recommends this model over d1-2B when there is. It says nothing about the 587M checkpoint's future, and Liquid AI has not deprecated it.

Running it, and what it costs in hardware

The inference path is Transformers 5.19 or later with trust_remote_code, and the call is model.system_one(state, questions, media). There are no sampling parameters to set, because nothing is sampled. Compilation takes about a minute on first calls; after that the card reports 5.8 ms for one question and 7.7 ms for three on an H100 PCIe, 3.7 ms and 5.6 ms on an AMD MI325X, and 21.1 ms and 28.9 ms on an Apple M5 Pro. A 5K-token state costs 52 ms on the H100. Batching is where the throughput is: 1,154 states per second on the H100 and 1,697 on the MI325X with 64 states in one call.

For a model in this class those numbers are the argument. It is not competing with a chat model on quality; it is competing with running a chat model at all. A 5.8 ms forward pass that returns a calibrated probability is a different line item from a generative call that returns a paragraph you then have to parse.

Where the decision ends and the rest of the pipeline starts

A model like this does one hop. It turns a ticket, a call recording or a camera frame into a label and a confidence, and then something has to act on that label — draft the reply, escalate the case, call the API. No Liquid AI model is in OrcaRouter's catalogue, so that second hop is where our product actually sits: the generative models that consume the decision go through one OpenAI-compatible key covering 200+ models, at each provider's list price passed through with 0% markup, with automatic failover when a request lands on a provider having a bad minute. The decision model is the cheap, fast, deterministic part. It is worth wiring the expensive, non-deterministic part to something that can retry.

What would settle it

Three things would move d1-Omni-2B from "promising" to "recommendable," and none of them requires another blog post. The first is an independent run of the audio suite on hardware the vendor does not control — the 22.8-versus-42.0 gap on spoken short states is the number most likely to move under someone else's microphone. The second is a published price for hosted access, without which the build-versus-buy calculation cannot be finished. The third is a direct comparison against d1-2B on the same audio-bearing tasks, because the card's own recommendation — use the 2B omni only when the input includes audio — is the most useful sentence in it, and it is also the one nobody has tested outside Liquid AI.