Daisuke Majima (MLBoy)'s banner

Daisuke Majima (MLBoy)

@JackdeS11 • 1,284 subscribers

400+ open models on Hugging Face, converted to LiteRT, Core AI and ExecuTorch and measured on phones. 163 of them in litert-community. [email protected]

Shorts

I converted Qwen-Image-2.1-Turbo to Core AI, Apple's on-device runtime. It is the 8-step version of the Qwen team's 7B text-to-image model. It can make transparent images. M4 Max: a 512×512 sticker with a real alpha channel in 9.7 s, timed by the app.

I converted Qwen-Image-2.1-Turbo to Core AI, Apple's on-device runtime. It is the 8-step version of the Qwen team's 7B text-to-image model. It can make transparent images. M4 Max: a 512×512 sticker with a real alpha channel in 9.7 s, timed by the app.

13,010 Aufrufe

GLiNER2.5-Decide (DeBERTa-v3-large, 340M) converted to LiteRT: fp16 weights, answers identical to the official fp32 model on 361/361 test requests, 66–70 ms per request on a Galaxy S26 GPU at 128 tokens, 175 ms at 256. Card, Python and Android sample:

GLiNER2.5-Decide (DeBERTa-v3-large, 340M) converted to LiteRT: fp16 weights, answers identical to the official fp32 model on 361/361 test requests, 66–70 ms per request on a Galaxy S26 GPU at 128 tokens, 175 ms at 256. Card, Python and Android sample:

15,814 Aufrufe