d1-omni vs Liquid AI D1 Ultra: An 850MB File Against a $0.042 Meter, and the Arithmetic Is Not What You Expect

The instinct when you put d1-omni next to Liquid AI D1 Ultra is that they occupy opposite ends of a range and the choice is about ambition. One is 587M parameters of open weights — 850MB on disk in the tag that appeared in Ollama's library this week, a checkpoint that has been downloaded about 13,900 times from Hugging Face since 5 October 2026. The other is Liquid AI's flagship decision model, announced on 15 October 2026, with no parameter count disclosed, no weights published, and a single published number that matters more than any benchmark: $0.042 per million input tokens. Renting a frontier decision versus owning a small one sounds like a budget question. Run the meter against a real workload and it mostly stops being one — the small model's advantage is not that it is cheap, it is that it is yours.

That distinction is worth pinning down precisely, because the two things being compared are not two models on the same shelf. Liquid AI D1 Ultra is an endpoint you call. d1-omni is a file you run. Both answer the same three question shapes — a yes/no with a probability, a label with a probability per label, or a position on a scale — and both return those answers in a single forward pass with zero output tokens generated. Everything after that, including what the answers cost and who is responsible when they drift, comes from which of the two you picked.

The rent side, priced the way Liquid AI actually bills it

Liquid AI D1 Ultra charges for input only, at $0.042 per million input tokens, with no completion charge because there is no completion. That is the whole rate card, and the interesting part is the image rule underneath it. Images bill at the text rate on a patch basis: 1.5 tokens per 32×32-pixel patch, which makes a 1024×1024 image 1,536 tokens. And every question is billed as its own prompt, carrying its own copy of the state and all attached images. Ask four questions about the same photo and you pay for that photo four times.

Put a number on a workload where a vision decision is the natural fit. Suppose a nightly pass over 25,000 product listings, each with one 1024×1024 photograph and about 200 tokens of listing text, asking two typed questions — one yes/no on whether the image matches the category, one score question on image quality. Each prompt carries 1,536 image tokens plus 200 text tokens, so 1,736 tokens, and there are two questions, so 3,472 tokens per listing. That is 86.8 million input tokens for the pass, or roughly $3.65. A nightly job, all month, lands in the low hundreds of dollars.

That is a genuinely uncomfortable result for anyone who assumed the frontier option was the expensive one. The pricing model is the reason: decision models never emit tokens, so the cost curve has no output term to blow up, and the per-question re-billing is a fixed multiplier you can plan around rather than a variable you discover in the bill.

The own side, and where its cost actually hides

d1-omni has no per-call price at all. It is a 587M-parameter model — 381M shared trunk and decision head, a 94M SigLIP2 vision encoder inherited from Liquid AI LFM2.5-VL-450M, and a 112M audio encoder — published under Liquid's LFM 1.0 terms, with GGUF builds following on 6 October and MLX builds on 12 October. Its cost is whatever hardware you already have, plus the part nobody puts in the comparison: the work of keeping it there.

What the two options do and do not share, in the dimensions that decide a deployment:

• Weights — published for d1-omni under LFM 1.0; not published for Liquid AI D1 Ultra. Liquid describes its flagship only as "based on a larger open-weight model with a series of Liquid-specific heavy customizations". The base is open; the model you call is not.

• Parameters — 587M and fully itemised for d1-omni; undisclosed for d1-ultra. You cannot size hardware for the flagship because Liquid has not published a size. You can size hardware for the 600M exactly: 850MB in Ollama's mixed Q8_0 tag, 1.2GB for the BF16 tag.

• Price — no per-call price on d1-omni; $0.042 per million input tokens on d1-ultra, with images at 1.5 tokens per 32×32 patch and every question re-billed with the full state.

• Modalities — d1-omni takes text plus images or text plus up to 30 seconds of audio, but never both in one request. d1-ultra is described as multimodal without a published modality breakdown; what Liquid documents is the question types and the image token rule.

• Context — 16,384 tokens for d1-omni across text, image and audio positions together, cut to 896 tokens of state and question text whenever an image is attached, "as trained". No context figure has been published for d1-ultra.

• Latency — no published inference numbers for d1-omni, because Liquid states it does not report them for a checkpoint it calls an early research release under active development; d1-ultra is served end-to-end with no separate latency figure published either.

• Calibration — the flagship was explicitly trained for it. Liquid's stated rationale is that most decision models are overconfident, so d1-ultra was tuned to produce more reliable probabilities. d1-omni carries a calibration tag of its own, but the 600M is an experimental checkpoint and no third party has measured either model's calibration.

The two scores are not one ruler

The temptation is to write 15.95 next to the flagship and subtract. Do not. d1-omni's 15.95 comes from Decision Index v0.2.1's public split, a board Liquid AI scored its own entries on with the official scorer rather than submitting them — sub-scores Knowledge 8.3, Language 12.9, Retrieval 35.0, Tools 15.1, Arts 6.8, against d1-3B's 48.57 inside the same family. Liquid's claim for d1-ultra is a first-place finish on the Artificial Analysis Decision Index by at least eight points, which is a different board with a different version, published before the flagship existed. The two numbers measure different things on different rulers, and the honest reading of the pair is only that the gap is large — not that it is 40-something points.

The figure that survives comparison is the one nobody publishes: the 600M has no inference numbers at all, so you cannot state its cost per decision on your own hardware without measuring it yourself. That measurement, not a leaderboard, is what makes the own-versus-rent call.

Three workloads, three different answers

• A nightly batch of 25,000 image checks — about $3.65 a pass on Liquid AI D1 Ultra, and on d1-omni a 1.2GB download plus whatever GPU you point at it. At this volume the meter is not the deciding factor; the deciding factor is whether you want the accuracy the flagship claims, and whether three dollars a night is worth more than maintaining a model server.

• A hard gate where the probability is the decision — an approval, a moderation ruling, a trade that fires on a threshold. Here calibration is the product, and d1-ultra is the only one of the two that Liquid says was trained specifically for it. A confident wrong probability at 0.6 is worse than no probability at all, and no third party has yet checked either model for it. If the threshold matters, you want the vendor's calibration claim reproduced before you trust it at the boundary.

• A high-volume, low-stakes classifier on a device — routing a voice command, triaging an inbox, flagging a listing for human review. Nothing about this workload wants a hosted frontier model. The 600M returns a typed answer from a single forward pass in 850MB, which is the entire reason it exists, and the 896-token image budget is fine for a sentence of state and one photo.

Where the routing boundary actually sits

Neither model is in our catalogue, and it is worth being blunt about that rather than implying otherwise: every Liquid AI d1-omni-600M call and every Liquid AI d1-ultra call goes to your own hardware or to the vendor's own API and third-party platforms, not through us. What OrcaRouter covers is the half of the pipeline a decision model cannot do for itself. A 587M decision model returns a label or a probability and stops; something still has to write the reply, summarise the case, or act on the score. One API for 200+ models, with each provider's list price passed through untouched and the routing DSL available when a triaged request needs to fan out to more than one model, is a quieter claim than "we host the decision model", and it is the true one.

What would change the answer

The comparison is currently decided by two absences. Liquid AI D1 Ultra has no published parameter count, so its cost per token cannot be reasoned about against the size of the job it is doing, and no latency figure, so an interactive path cannot be designed around it without measuring. And d1-omni has no published inference numbers by the vendor's own choice, so the own-versus-rent calculation has one side that only a benchmark of your own hardware can fill in.

If either of those closes, the shape of the decision changes. An independent calibration result on d1-ultra would settle the gate workload, where the price is already an order of magnitude cheaper than most people would guess. A latency figure from Liquid for the 600M would let someone place it in an interactive path with confidence instead of testing it first. Until then the safe reading is narrow and unglamorous: the flagship is priced close enough to free that budget alone will not push you toward running anything yourself, and the 600M's real advantage is that once you have the file, nobody can change the number, the terms, or the endpoint under you.

The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.