Download

llm-hardware-compatibility-2026-10-03-13163f6c89cd.csv

SHA-256

13163f6c89cd8a2fa548c9c062d9d1e9d6d878d30f0545f298769853b4cc61b3

Format · text/csvSize · 11.77 MB

llm-hardware-compatibility-2026-10-03-13163f6c89cd.json

SHA-256

23e16051117f674949037a706f514ced22b402a9dfd05a773b7002c49db5d991

Format · application/jsonSize · 36.23 MB

The checksums are the point of publishing them: a file that has been edited between here and wherever you found it will not match. manifest.json carries the same figures for a machine to read.

What is in it

One row for every catalogued model and device: 44,145 rows for 327 models and 135 devices, sized at 8,192 tokens of context with the whole model resident and offload switched off — the same question the pairing pages answer, computed by the same engine. A device sold in several memory configurations is sized at its largest.

297 models are analysed (status ok).
23 models publish a maximum context below 8,192 tokens; their rows say so (status context_exceeds_model_max) rather than being resized to a different workload.
7 models have no usable size source; their rows say so (status not_sizable). The coverage page says which and why.

Version 2026-10-03-13163f6c89cd is the build date and the first twelve characters of the CSV’s SHA-256. Snapshot ZOAGzSDaZvVT, engine 2026-10-04.1. A refresh publishes a new, differently named version and never rewrites this one; earlier versions stay downloadable and are listed in the manifest’s history.

The columns

Column	What it holds
`model_slug`	Catalogue identifier of the model, and the last path segment of its page.
`model_name`	Name as its publisher writes it.
`publisher`	Namespace the configuration was read from (for a mirror, the mirror's namespace).
`hf_repo`	Hugging Face repository the architecture was read from.
`config_revision`	Immutable commit of that repository, so the row can be reproduced.
`parameters`	Parameter count from the published safetensors metadata; empty when unusable.
`device_id`	Catalogue identifier of the device, and the last path segment of its page.
`device_name`	Product name as its manufacturer writes it.
`vendor`	Manufacturer of the device: NVIDIA, AMD, Apple or Intel.
`device_memory_gb`	Memory the manufacturer publishes, counted as GB × 10⁹ bytes. For a device sold in several configurations, the largest.
`device_bandwidth_gb_s`	Published memory bandwidth of that configuration.
`bandwidth_evidence`	verified when the manufacturer publishes the bandwidth; derived when it is arithmetic over a published bus width and data rate.
`context_tokens`	Context every row is sized at (8192).
`status`	ok: analysed. context_exceeds_model_max: the model's published maximum is below the reference context. not_sizable: no usable size source for this model. unsupported: the engine refused, see status_reason.
`status_reason`	Why a row is not ok; empty when ok.
`fits`	For status ok: true when at least one format fits with the whole model resident and no offload. Empty otherwise.
`best_format`	Highest-quality format that fits, empty when none does.
`required_bytes`	Memory that format needs: weights, KV cache and runtime reserve.
`utilisation`	required_bytes divided by device memory, at the best format that fits, or at the smallest format when none does.
`weight_size_evidence`	verified: a published weight file; derived: reconstructed from the architecture; bounded: a range from the parameter count.
`smallest_format`	Smallest format the model ships in, which is what a 'no' is measured against.
`shortfall_bytes`	How far the smallest format misses by, empty when something fits.

Licence

The compilation and the computed columns are published under CC BY 4.0: use them anywhere, including commercially, with attribution. Attribution is a condition of the licence for the dataset; for a single figure quoted from a page it is asked for, not required.

Shipping the dataset inside a product where a credit line does not fit — an app’s model picker, a marketplace listing? The data licence covers exactly that: the same files, no attribution requirement, and a Business key for everything the files do not precompute.

The facts underneath are a different matter, and this is not a lawyerly distinction. A manufacturer’s published bandwidth and a publisher’s configuration file are not ours to license, and claiming otherwise would be asserting ownership of somebody else’s specification sheet — which is what the terms have said since before this file existed. What is licensed here is the work that is genuinely ours: gathering those figures, computing the answer, and labelling how strong each piece of evidence is.

One condition worth stating plainly, because it is the failure this site exists to avoid: the weight_size_evidence and bandwidth_evidence columns are part of the data, not decoration. Republishing a derived figure as though it were measured misrepresents the work rather than merely reusing it.

Citing it

LLM Bottleneck open compatibility dataset (llmbottleneck.com/data), version 2026-10-03-13163f6c89cd, CC BY 4.0. https://llmbottleneck.com/datasets/llm-hardware-compatibility-2026-10-03-13163f6c89cd.csv

Applies to the compilation, the computed columns and the evidence labelling only. Model weights, model cards, configurations, product names, trademarks and manufacturers' specifications belong to their owners and are not relicensed; they are cited by source URL.

For a single figure rather than the whole table, the citation page explains what to cite us for and what to cite the publisher for, and the API answers one configuration at a time.