Liu Liu
@liuliu
Maintains / / Wrote iOS app Snapchat (2014-2020). Founded Facebook Videos with others (2013). Sometimes writes at
284 Following 4.2K Followers
I don’t know why this needs to be re-iterated: if you are lucky and having a M5 Max / M5 Ultra, don’t use Qwen 3.8 27B, you will be much happier with a low-bit DeepSeek 4.1 Flash 1017. A good 2-bit quant for that model is just so much better than 27B dense.
Show more
The best way to run DeepSeek 4.1 Flash 1017 on M5 Ultra is to use @localcodeapp. It is fully open-source: it is usable, and it is fast (note the decode can be improved further with DSpark).
Show more
Follow-up on yesterday's hot-take: there are a lot of noise, and you should be suspicious of any claims such as 4x, 8x, 10x speed-up over @awnihannun's MLX or @ggerganov's llama.cpp with "custom / model-specific inference engines" on Apple's hardware.
Show more
Just for the fun. Note that our default is the "Max Quality" in some other apps w/ Apple Silicon. You have to enable TeaCache to get explicit "approximate quality".
