4 ms·
Show HN: Vram.run – Compare API providers, local GPUs, and cloud for any model
- jad-nohra 8mo ago[dead]
- ranger_danger 8mo agoThe data for Llama-4-Scout seem quite wrong to me... it says a 17B model is only 7B, and that Q4 is 4GB when it's more like 60GB.
- jad-nohra 8mo agoThank you I will take a look at that. It looks like MoE models like Llama-4-Scout report active params (17B) not total (~109B). Some MoE models already work correctly when HuggingFace has the full metadata. Fix is in progress. Fixed in v0.3.1 and is live. https://vram.run/model/meta-llama/Llama-4-Scout-17B-16E-Instruct/ https://vram.run/model/meta-llama/Llama-4-Scout-17B-16E-Inst...