Models chosen to fit real consumer GPUs.
Small, open-weight, quantised — because a marketplace on idle gaming hardware only works if the models actually fit in that hardware.
Launch candidates
DEMO — illustrative only, not real data
| Model | Licence | Size (Q4) | VRAM needed | Status |
|---|---|---|---|---|
| Qwen3-4B | Apache 2.0 | ~2.5 GB | ~4 GB | Planned — not yet serving live traffic |
| Qwen3-8B | Apache 2.0 | ~5.2 GB | 6–8 GB | Planned — not yet serving live traffic |
| Qwen3-14B | Apache 2.0 | ~9.3 GB | ~12 GB | Planned — not yet serving live traffic |
| Llama-3.1-8B | Llama 3 Community License — commercial use under 700M monthly active users | ~5 GB | 6–8 GB | Planned — not yet serving live traffic |
| Mistral-7B | Apache 2.0 | ~4.5 GB | 6 GB | Planned — not yet serving live traffic |
Sizes are approximate Q4_K_M quantised weights; VRAM figures include runtime overhead. Model identifiers in the API: qwen3-4b, qwen3-8b, qwen3-14b, llama-3.1-8b, mistral-7b. Licences were checked against the model publishers’ published terms at research time; final serving depends on a licence re-check before launch.
How models are chosen
Small (1B–4B) first for broad compatibility, then medium (7B–8B), then larger (12B–14B) for machines with 12 GB+ VRAM. A model is only listed as supported once it has actually run and been benchmarked on representative provider hardware — never on a spec sheet alone.
Chat playground
Talk to the demo backend. This POSTs to http://localhost:8787/v1/chat/completions with an OpenAI-compatible body. If the backend is not running, the error is shown honestly below — nothing is faked.