README.md

---
pretty_name: Self-Hosted AI — GPU Compatibility, Recipes and Catalogue
license: cc-by-sa-4.0
language:
  - en
size_categories:
  - 1K<n<10K
task_categories:
  - table-question-answering
  - text-retrieval
tags:
  - gpu
  - vram
  - local-llm
  - quantization
  - hardware-compatibility
  - self-hosted
configs:
  - config_name: compatibility
    default: true
    data_files: data/compatibility.parquet
  - config_name: recipes
    data_files: data/recipes.parquet
  - config_name: models
    data_files: data/models.parquet
  - config_name: gpus
    data_files: data/gpus.parquet
  - config_name: benchmark_sources
    data_files: data/benchmark_sources.parquet
---

# Self-Hosted AI — GPU Compatibility, Recipes and Catalogue

Which open-weight AI models actually run on which consumer GPU, and what it takes
to get them running. 2 700 model×GPU verdicts across 100 models
and 27 cards, plus 1 009 full setup guides
(19 MB of markdown) written against specific hardware.

This is the machine-readable form of [smeltcore.com](https://smeltcore.com). Every row carries a
`url` back to the page it came from.

Generated 2026-09-24T19:45:21+00:00 from the public read API (`https://api.smeltcore.com/api/v1`) — no private data, no
credentials, reproducible by anyone.

## Configs

| config | rows | what it is |
|---|---:|---|
| `compatibility` | 2 700 | **the point of the dataset.** One row per model × GPU, with a verdict |
| `recipes` | 1 009 | full setup guides, markdown included, tagged by model / GPU / tool |
| `models` | 100 | the catalogue: licence, upstream repo, modality |
| `gpus` | 27 | the cards: VRAM, vendor, series |
| `benchmark_sources` | 166 | third-party measurements, normalised and cited |

```python
from datasets import load_dataset

compat = load_dataset("REPO_ID", "compatibility", split="train")
compat.filter(lambda r: r["gpu_slug"] == "rtx-4090" and r["fit"] == "verified")
```

## The `fit` scale

The whole dataset turns on this column, so it is worth reading before using it.

| verdict | rows | meaning |
|---|---:|---|
| `verified` | 1 008 | somebody ran it on this exact card and wrote down how — there is a recipe behind the row |
| `fits` | 659 | inferred: the model's memory floor is under the card's VRAM, and the vendor is supported. Not measured |
| `unknown` | 461 | no floor established for this model, so no honest call can be made |
| `too_big` | 572 | the memory floor exceeds this card. This is the one verdict asserted from anywhere, not only from same-vendor evidence |

The asymmetry between `fits` and `too_big` is deliberate: a model is only called
runnable on evidence from the same vendor's hardware, but it is called *not*
runnable from a memory floor established anywhere. Being wrong in the optimistic
direction wastes somebody's evening; being wrong in the pessimistic direction
only costs them a model they could have tried.

`min_vram_gb` is a **filter floor in decimal GB** — the smallest card the model is
offered on — not a measured peak. Measured peaks, where they exist, are in
`peak_vram_gb` and in `benchmark_sources`.

## Provenance, stated plainly

**`recipes` is first-party.** Written for this catalogue against named hardware,
with the quantization, runtime and settings each one was written for.

**`benchmark_sources` is not.** 111 of 166 rows come from a
single third-party site (`www.hardware-corner.net`); 9 were measured by us, each linking to its
raw session in [Smeltcore/measurements](https://huggingface.co/datasets/Smeltcore/measurements), and
3 were submitted by readers through the site.
It is published as a **citation index, not as our benchmarks**: what this project
contributes is the normalisation — one model slug, one GPU slug, one unit
convention — and every row is required to carry `source_url` back to whoever did
the measuring. Credit and verification both belong there. If you use a number from
this table, cite the source row, not this dataset.

`confidence` is a 0–1 score reflecting how much the source is trusted; it is not a
statistical confidence interval.

## Coverage and what it is not

- 27 consumer cards — NVIDIA, AMD and Apple silicon. **No datacenter GPUs**
  (no H100, no A100): this catalogue is about hardware people own.
- 8 modalities: llm (38), multimodal (18), image (15), video (14), tts (6), 3d (4), music (3), specialized (2).
- Verdicts are about **whether it runs**, not how well it performs. There is no
  quality benchmark here and no leaderboard.
- The catalogue moves — models get added, quantizations appear weekly. A stale copy
  of this dataset will understate coverage. `generated_at` above is the only date
  that matters.

## Licence and attribution

Released under **CC BY-SA 4.0**, matching the licence the site publishes its data
under. Attribution goes to [smeltcore.com](https://smeltcore.com).

Rows in `benchmark_sources` describe third-party work; that licence does not
extend to the measurements themselves, which belong to the sites named in
`source_url`.

## Citation

```bibtex
@misc{smeltcore_selfhosted_ai,
  title  = {Self-Hosted AI — GPU Compatibility, Recipes and Catalogue},
  author = {smeltcore},
  url    = {https://smeltcore.com},
  note   = {Generated 2026-09-24T19:45:21+00:00},
  year   = {2026}
}
```