d1-Omni-2B vs d1-omni: One Tag, Two Checkpoints, and an Eightfold Gap

Pull d1-omni today and you get a 587M-parameter decision model in 850MB. Pull d1-omni:2b and you get 4.4GB of something else entirely — d1-Omni-2B, the audio decision model Liquid AI shipped with the LFM3 family on 27 October 2026, against d1-omni, the name Ollama has used since early October for the vendor's d1-omni-600M checkpoint. They are not two vendors' answers to the same question. They are two generations of the same idea, eight times apart in parameter count, and the choice between them is really a choice about where the model has to run. One fits on a phone. The other wants a GPU and, in exchange, does things the smaller one cannot do at all.

The naming makes this harder than it should be, so it is worth stating plainly before the numbers: there is no Liquid AI model called simply d1-omni. The vendor's documentation lists d1-omni-2B and d1-omni-600M as open weights, and d1.1 and d1-ultra as hosted-only. "d1-omni" is a packaging name — the 587M checkpoint under a shorter label. That is the opponent in this comparison, and it is the one most people already have on disk.

What each one actually is

Both are decision models in Liquid AI's sense of the term: you hand them a state and a list of named questions, and they return typed answers with probabilities in a single forward pass. Neither writes text, neither produces speech, and both report zero output tokens. The similarity ends at the interface.

• Parameters — 2.17B for d1-Omni-2B against 587M for d1-omni, split in the small model as a 381M shared trunk and decision head, a 94M vision encoder and a 112M audio encoder.

• Architecture — an LFM3 tri-hybrid backbone for the 2B: 24 layers of 13 gated short convolutions, 5 sliding-window attention and 6 global attention. The 587M is an LFM2.5 bidirectional encoder built on LFM2.5-Encoder-350M, with a SigLIP2 vision tower inherited from LFM2.5-VL-450M and a 17-layer FastConformer for audio.

• Context — 128K tokens with a 128,000-token vocabulary for the 2B. The 587M model is documented at 16,384 tokens covering text, image and audio positions together, and its card adds that when images are present the state and question text is cut to 896 tokens, as trained. Ollama's page prints 125K for both tags, a figure the vendor does not publish; size your prompts against the vendor's number.

• Input — the 2B takes a text or JSON state plus pictures, video frames and audio clips in one media list. The 587M takes text with images, or text with up to 30 seconds of audio, and its card does not document mixing the two.

• Output types — yes/no, choice and score on both. The 2B adds boxes, multi-select, extracted values with spans, and numbers with an interval. Those are new in the LFM3 generation and have no equivalent on the 587M.

• On disk — 4.33GB of BF16 safetensors for the 2B, against 850MB for the Q8_0 build of d1-omni and 1.2GB for its BF16 tag.

• License — LFM Open License v1.0 on both. Free for commercial use below $10,000,000 in annual revenue, separately licensed above it.

What the eighth parameter buys

The useful comparison is the one Liquid AI printed itself, because it re-scored all four d1 models on the same items with the same scorer. That makes the rows internally comparable in a way the older cards are not.

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

• DecisionBench (Hanno Labs) v1 — 76.4 against 58.7.

• decisionbench.ai v4 — 84.7 against 57.9.

• JevBench v1.4.1 — 75.3 against 58.0.

• ImajevBench v2.0-lite — 72.3 against 21.7, the widest gap in the set and the one that matters most for visual work.

• WorkflowEvals — 77.1 against 48.8.

• Typed Decisions — 63.1 against 57.5, the narrowest gap and a reminder that not every task scales with parameters.

• Vision average, 18 image benchmarks — 67.7 against 42.8.

Two cautions travel with that list. The first is version drift: the 7 October card reported d1-omni-600M at 15.95 on Decision Index v0.2.1, and the new card reports 18.3 on v0.3. Those are different suites, and subtracting them measures the index, not the model. The second is provenance: every figure is vendor-reported, scored by Liquid AI, on benchmarks Liquid AI selected. Nobody outside the company has published a run of either checkpoint.

The shape of the result is consistent, though. On classification-style tasks the 587M holds up surprisingly well — Typed Decisions is within six points. On anything that requires reading a scene, counting objects or interpreting a layout, it collapses: 21.7 on ImajevBench against 72.3, and 14.9 on VisualWebBench in the earlier suite against 88.6. If your job is routing a text ticket, the extra parameters are a luxury. If your job involves looking at something, they are the whole product.

Where the 587M still wins, and the number that would prove it

Deployment envelope is the 587M model's real advantage, and it is not a consolation prize. A 587M model in 850MB runs on a laptop, a Raspberry Pi or a phone, and Liquid AI built the d1 line for exactly that. A 2.17B model in 4.33GB needs a GPU or a well-provisioned Apple Silicon machine before it is pleasant to use.

What is missing is the latency comparison that would make the trade precise. The d1-Omni-2B card publishes a full speed table: 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 MI325X, 21.1 ms and 28.9 ms on an M5 Pro, and 1,154 states per second batched on the H100. The d1-omni-600M card publishes nothing, and says why: "We don't report inference numbers for d1-omni-600M as it is an early research release and is under active development." So the smaller model is almost certainly faster, and there is no vendor figure to quote for how much. Anyone who tells you the 587M is three times faster is extrapolating from parameter count.

There is one genuine accuracy-adjacent advantage worth naming. The d1-Omni-2B card recommends d1-2B over itself when there is no audio, and gives the reasons: d1-2B is stronger on text and single boxes, reports missing values more often, and is better calibrated. The same advice points away from the omni model for text-only work. It does not say the 587M is better calibrated — only that the 2B omni carries a cost for a capability you may not use.

The audio split is where the two generations separate cleanly

Audio is the only modality where both models compete directly, and the constraints are different enough to decide an architecture.

The 587M model takes up to 30 seconds of speech, expects 16kHz mono, and its card carries an explicit warning: audio capabilities were trained on requests between an English speaker and an assistant, covering what kind of utterance it is, its topic and what the speaker wants. That is a narrow distribution, and it is stated as a limitation rather than a footnote.

The 2B model supports speech in the same sixteen languages as its text side, minus Hindi, Thai, Vietnamese and Indonesian, and the card's audio results include 84.5 on ESC-50 sound events and 92.5 on FLEURS language identification across 43 languages. It also degrades in ways the card documents: spoken short states drop from 42.0 written to 22.8 heard, and Belebele passages in eight languages drop from 83.6 to 41.8. Long spoken passages lose accuracy, most of all outside Latin script.

So the split is not "the 2B is better at audio." It is that the 2B hears more kinds of audio and still struggles with long spoken state, while the 587M hears a narrower, more controlled slice and was never benchmarked on sound events or language identification at all. For voice-command routing in English, the 587M's constraints are survivable. For multilingual call triage, they are disqualifying.

Which tag to pull

The decision rules are short enough to state as rules.

• Text and images only, and no audio in the product — pull neither. d1-2B scores higher than d1-Omni-2B on the decision index, 51.1 against 49.0, and the vendor's own card tells you to use it. The omni premium buys audio you will not use.

• Voice routing on constrained hardware — d1-omni at 850MB. Accept the 30-second cap, the English-speaker training distribution and the 16,384-token context, and measure your own latency because the vendor will not give you a number.

• Audio plus images in the same decision, or boxes and spans in the output — d1-Omni-2B. The 587M cannot express those requests, so there is no comparison to make.

• Multilingual speech — d1-Omni-2B, with the four excluded languages checked against your actual traffic before you commit.

• Batch throughput — d1-Omni-2B, whose card is the only one of the two that publishes states per second at all.

One adoption signal is worth recording because it cuts against the benchmark story: d1-omni-600M has drawn roughly 18,900 downloads and 148 likes on Hugging Face since 5 October. d1-Omni-2B had 36 downloads and 7 likes four days after its repository appeared. Early download counts measure novelty and packaging, not quality — the 587M model has an Ollama tag that makes it one command away — but they do describe which model people are actually running today.

After the decision, the rest of the call

Whichever tag you pull, the decision model is the first hop and not the last. It returns a label with a probability; something still has to write the reply, escalate the case or call the next API, and that second hop is a generative one. OrcaRouter does not carry Liquid AI models, so the d1-omni or d1-Omni-2B call runs where you already run it — on your own hardware from the Hugging Face weights, or through the vendor. What sits behind one OpenAI-compatible key is the other 200+ models in the catalogue, at each provider's list price with 0% markup, with failover across providers when the downstream model you chose is the one having a bad afternoon. For a model this new, that separation is useful: the unproven part stays local and cheap, and the part that touches production traffic gets a retry path.

The number neither card gives you

Neither checkpoint has an independent evaluation. The 587M has been public since 5 October and still has none; the 2B is four days old. Until someone runs both on the same audio set with the same microphone and the same languages, the eightfold parameter gap is a vendor's table, not a measured fact. The specific figure to watch is that 22.8 on spoken short states — it is the one row where the new model's own card admits a weakness large enough to change a deployment decision, and it is the row most likely to move when somebody outside Liquid AI points a microphone at it.