Which releases name consumer hardware at all
Three of the 15 families name a consumer card in their hardware statement. Five name data-centre cards or chip architectures, four state gigabytes with no hardware named, and three state nothing at all. As of 2026-09-12.
| Item | What is published | Detail |
|---|---|---|
| A consumer card named | Three families | A 4090 with a timing and an output; one RTX 4090 for a 4.5B model; RTX 30XX to 50XX with the formats |
| Data-centre cards named | Three families | At least one H100; H100 or H800 in counts of one, four and eight; an A100 80GB |
| A chip architecture named | One family | Ampere, Hopper and Blackwell |
| A card size and no model | One family | A single 80G GPU, behind two peak figures |
| Gigabytes and no hardware | Four families | A figure with nothing named beside it |
| Nothing stated | Three families | No figure and no card |
Inclusion rule. One row per class of hardware statement. A row is kept where one family uses that form, because the class of machine named is the audience signal. Order. From consumer hardware outward to no statement at all.
1A consumer card claim is falsifiable, which is the point
One repository names a specific consumer card, the release it applies to, the output produced and a duration. Every part of that can be checked by somebody who owns the card, which is what separates it from a general claim about local running.
Most consumer-hardware claims in this field name neither the release nor the output. Those are unfalsifiable and therefore tell a reader nothing.
2Naming untested generations is worth more than a blanket claim
One repository lists the card generations it supports and says two older generations are not tested, rather than calling them unsupported. That is a smaller claim and a more accurate one, and the register keeps the wording.
It also names the numeric formats, which is the part a memory figure cannot carry. A card with the capacity and without the format does not qualify.
3The two classes do not convert
A figure measured on a data-centre card does not transfer to a desktop one, because the formats, the bandwidth and the driver stack differ. A figure measured on a consumer card does not describe a cluster either.
So the register records the class the vendor named and leaves the rows uneven. A normalised column would have required inventing the other half of every statement.
- HardwareThe repository states TI2V-5B runs on consumer-grade graphics cards such as a 4090, generating a 5-second 720P video in under 9 minutes on a single consumer GPUvendor's own figure
- HardwareH100/H800 x 8 for the 24B model, RTX 4090 x 1 for the 4.5Bstated as a count of a named card
- HardwareNvidia GPU in RTX 30XX, 40XX, 50XX series that supports fp16 and bf16; the GTX 10XX/20XX are not testedcard generations and numeric formats
- HardwareApproximately 60GB VRAM on a single GPU, with at least one H100 recommendedthe vendor's own figure
- HardwareThis model requires 32.54 GB of GPU VRAM, on Ampere, Hopper or Blackwell hardwarea figure and the supported architectures
4Sources
Repository statements and hosted rates are linked from the model pages and the how read page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: Stated settings, Card or repository, More than one card.