AI
LLM Services
Approved Peft members can experiment with private language models for game dialogue, stories, and structured text, without supplying an API key or paying per token. These services support only online/HTML games that do not require Windows to play. Requests share a bounded queue and may wait while a model loads; keep one model per game where possible to avoid repeated switching. Native reasoning stays enabled where the model uses it, so there may be a pause before the visible reply. Peft does not save this test conversation: history stays in this page and is sent to the private model service with each turn. Do not enter secrets or sensitive personal information.
Observed response times
Same short NPC prompt; one completed request per model (39–50 visible words). Native reasoning included, not displayed. Recorded 2026-11-10 (UTC), not a speed or quality guarantee. Benchmark models may not all be offered in the tester.
| Model | First visible text | Full reply |
|---|---|---|
| FastGemma 4 12B | 0.25 | 1.68 |
| DefaultGemma 4 26B-A4B | 0.38 | 1.34 |
| QualityGemma 4 31B | 117.32 | 124.67 |
| StrataQwen3.8 Flash Next | 2.49 | 2.82 |
Model switching costs
Average seconds to the next model's ready response, including stopping/loading. Two trials per direction on 2026-11-10. Queue waits and answer generation are extra.
| From ↓ / To → | Fast | Default | Quality | Strata |
|---|---|---|---|---|
| Fast | — | 13.8 | 10.7 | 21.7 |
| Default | 6.9 | — | 11.8 | 25.3 |
| Quality | 6.9 | 10.9 | — | 22.4 |
| Strata | 9.7 | 16.1 | 13.0 | — |
Try a model
Sign in with an approved, active Peft membership to test the models.
Sign inPrompts
Give your game's coding assistant a prompt to add one Peft feature, or combine the features you missed.
Check prerequisites before selecting. Already implemented them? Select only what is missing. Composition includes exactly your selections, in prerequisite-first order.