jackailocal / docs /HACKATHON_MODEL_SELECTION.md
jackboy70's picture
Deploy: accurate lite-builder note
f25362a
|
Raw
History Blame Contribute Delete
1.02 kB
# Hackathon Model Selection
The hackathon package should demonstrate that JackAILocal can choose models up to the allowed small-model limit while still offering a safe default.
## Default
Use a 4B or 9B model for the public demo because it is practical on normal hardware.
## Power profile
When the uploaded, manually entered, or locally scanned profile shows 24 GB VRAM, unlock the Power profile and recommend a 27B or 30B candidate that stays inside the 32B rule.
## Rule
Do not include models above the hackathon maximum. Do not include ambiguous models whose published parameter count exceeds the rule.
## Qwen3.6-27B update
Qwen3.6-27B is allowed as a Power profile candidate because it is a 27B model and stays under the hackathon 32B cap. It should be presented as a high-end profile for RTX 3090-class hardware, not as the ordinary laptop default.
The default demo should still run with a smaller model, such as Qwen3.5 4B or 9B, because the Space and common laptops should not depend on 24 GB VRAM.