MeteorLLM 0.3.0 Brings Qwen3.5-0.8B to PS Vita
MeteorLLM v0.3.0 brings local Qwen3.5-0.8B inference to the PlayStation Vita, with runtime close to the previous 0.6B build.

MeteorLLM, an experimental offline language model app for the PlayStation Vita (PS Vita), has updated to version 0.3.0. The new release replaces the older Qwen3-0.6B engine with Qwen3.5-0.8B.
The bigger model runs at close to the old one’s speed. Dropping some matrix weights from Q6 to Q4 frees enough RAM to cache frequently used output weights. That cuts storage reads during generation.
In testing with a short prompt, v0.3.0 finished in 34.32 seconds. That is a little faster than the older 0.6B build’s 37.14 seconds on the same question, but these are separate single runs, not a benchmark. Everything runs locally and offline, and the full chat UI fit with about 5 MiB of system RAM free in the tested cases. The model still makes factual and arithmetic mistakes, and the timings came from a test device with an external overclock set to 500 MHz.
How to try it
- Install the MeteorLLM-v0.3.0.vpk file.
- Download
MeteorLLM-Qwen35-0.8B-model.zipfrom the release page. - Extract the four model files directly into
ux0:data/VitaDesk/MeteorLLM-08/on your memory card.
The old 0.6B files are incompatible with this engine, but they stay untouched in their old folder, so you can reinstall v0.2.1 to roll back.
