Piotr Mazurek (in Warsaw 🇵🇱)
@tugot17
948 Following 4.4K Followers
highlight of my Warsawmaxxing 🇵🇱 was discussing local AI with @Prince_Canuma 🫡
ok, so it's not sharing across layers like in 5.2, but pooling across neighbouring tokens
1200GB/s of memory bandwidth - 4.5x the bandwidth of DGX Spark 🍏🫡
benchmarking on real hardware is infamously challenging; amazing release by my colleagues at @liquidai and @ArtificialAnlys 🫡
*yes upper and left is better😅
Ok now I’m MLX-VLM pilled 😅
Big conviction bet on behalf of Stripe; they have to believe they can scale Open Router by 3 orders of magnitude; the daily liquidity of the most popular model (DSV4 Flash) is ~$30k in output tokens (input hard to estimate cause of caching)
Show more
