Pairs whose published answers point opposite ways

A pair gets a page when two families state something in the same column and the two statements would send a reader to different hardware, different terms or a different plan. Both values are quoted from the vendors' own pages. As of 2026-09-22.

Fig. 1 Silence on one side counts: an exact figure beside a blank at a comparable model size is an asymmetry worth a page.

Most pairs in a register agree, and agreement is dull to read twice. Two releases of similar size that both state a memory figure in the fifties tell a reader one thing, and the column summary already tells it. The pairs collected here are the ones where the same question gets answers an order of magnitude apart, or where one vendor answers and the other declines.

The pairing is never a ranking. A release that states a low memory floor is not better than one that states a high one; it is a different model, trained differently, documented for a different reader. What the pair shows is how far apart two published answers to one question can sit, which is the thing a plan has to survive.

Silence counts as an answer here. Where one family publishes an exact figure and another publishes none for the same column at a comparable model size, the asymmetry is the finding, and it is recorded as an absence rather than filled in from a forum thread.

1Opposite answers in the licence column

Bespoke or MIT — Two large releases, one under a company's own licence name and the other under MIT, which is the difference between reading a document and recognising one.
Split, or one document — One family licenses its 2B and 5B releases differently; the other applies a single company licence across 1.3B, 5B and 14B variants.
Borrowed or authored — One card carries a licence named after a company that did not publish the weights; the other was written by its publisher and answers commercial use.
Pointer or answer — Two vendor-written licences and two approaches to the commercial question: one directs it to a separate page, the other settles it on the card.
Apache or bespoke — Two releases of comparable size, one under Apache 2.0 and the other under terms its vendor wrote, which decides how much reading precedes a download.

2Opposite answers in the size column

Exact against a bound — One card publishes 2,059,174,912 parameters and the other publishes over thirteen billion, which are two different kinds of claim about a file.
Decomposed or round — One card publishes 175M and 2.8B as separate components; the other publishes thirty billion parameters as a single description.
No count, or three — One card publishes no size at all while documenting its outputs; the other publishes 1.3B, 5B and 14B with memory figures for two of them.
Count and cost inverted — The larger count comes with the smaller memory floor, which is the clearest case in the register against reasoning from parameters to hardware.

3Opposite answers in the hardware column

6GB against 78GB — Two published memory statements thirteen times apart, one a floor for a minute of video and the other a peak that never drops below seventy gigabytes.
Same size, one answer — Almost identical parameter counts, and one repository publishes a floor with the cards named while the other publishes nothing at all about the machine.
Floor or builds — Both releases are 13B and both address memory: one with a stated minimum and card list, the other by publishing distilled and fp8 variants instead.
Two 2B, one figure — Both releases are around two billion parameters.
Silence against decimals — One card names the memory technique and never gives a number; the other gives a figure to two decimal places and names three GPU architectures.
4GB against 60GB — The lowest and one of the highest published floors in the register, one stated in FP16 for a 2B model and the other for a 10B model at 480p.
Cards or gigabytes — Both releases describe multi-GPU running.
Two pairs, two variables — Both repositories give two memory figures.
Both sides, or neither — Allegro states memory with and without offloading for the same run; LongCat-Video states no memory figure at all for a model four times its size.

4Opposite answers in the resolution column

Hedged or specified — One repository hedges its output size as a current capability; the other publishes size, frame rate and clip length in a single sentence.
Borrowed or published — One repository's only pixel figures are the settings of its memory table; the other publishes a full output row with frames, seconds and rate.

5Opposite answers in the length column

Four seconds or sixty — The shortest and the longest published lengths in the register, fifteen times apart, one written as a limitation and the other as a supported ceiling.
Frames, or frames and rate — One repository caps at 204 frames with no frame rate anywhere; the other publishes frames, seconds and rate together so nothing needs converting.
A table or no ceiling — One repository maps frame counts to durations at four points; the other treats frames as an argument and states no maximum at all.
Priced, or unstated — One repository states a minute of output and the memory it needs in one sentence; the other states no clip length anywhere.

Single values, one family and one column at a time, are on statements. The column itself is described on fields.

Also in the register: Models, Hosted, Statements, Fields, Questions, Terms, Learn, Data. What a release has to publish is set out on the reading page.