Make AOT compilation conditional for models >= 2B parameters to optimize free tier usage

Add AOT compilation optimization for ZeroGPU acceleration

add 4 20b+ models after enabling dynamic gpu duration

Add dynamic duration calculation for ZeroGPU acceleration

make qwen-4b default

disable two models that cannot run or too run too slowly on hf spaces with zerogpu

Update app.py

Update app.py

Update app.py

add original apriel 15b

use apriel 8bit

run Apriel on 4bit

Add Apriel-1.5-15b-Thinker

Update app.py

Update app.py

Update requirements.txt

Update requirements.txt

Update requirements.txt

Update app.py

add 4 models from qwen3 family

add qwen3 32b awq

Update app.py

Update README.md

Update requirements.txt

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

feat(models): add Granite-4.0-Micro and Qwen3-4B-Instruct-2507 to MODELS registry

feat(models): Added three new models

add 5 models from liquid ai

add smollm2 135m multilingual

add parser_model_ner_gemma_v0 based on gemma 3 370m it

Add Gemma-3-270m Taiwan

add gemma-3-270m-it

remove prevously added breeze models (as it didn't work), add smollm 135m taiwan

add breeze models

add 3 sub-1B TW models from ShengweiPeng

add Qwen2.5-Taiwan-3B-Reason-GRPO & Llama-3.2-Taiwan-1B

disable cache for better compability

make qwen2.5-taiwan 1.5b the default

upgrade transforemrs to 4.53.3+

upgrade transformers to 4.46.1+

add timm to requirements

add access_token to model loader

add gemma 3n, smallthinker 4b0.6b and falcon h1-1.5b

rename duckduckgo_search to ddgs

remove qwen3 30ba3b & 32b, but add 14b

use fp8 version of qwen3 30ba3b & 32b