Really impressed about the Microsoft Mage-VL: a 4B vision-language model that runs on your gaming PC (~10GB, any 16GB GPU, Apache 2.0)
it's codec-native, follows how video compression works to cut 75% of visual tokens, 3.5x faster. plus a streaming gate that watches video and only speaks when something happens
Event-gated video understanding on one consumer GPU. That used to be cloud-only
https://t.co/gsgPj4ki7u
About Kimi K3 recently released
The point is that hosting frontier intelligence is no longer a permission slip issued by three labs.
Any cloud, any country, any startup can serve it now.
@TheAhmadOsman storage: ~2TB, easy
running: the HF repo is 1.56TB. vLLM's metadata says 1,680GB min VRAM, and that's before KV cache, comms buffers, vision encoder, and concurrency
Like before, we used to be impressed with 2GB of RAM perhaps 1,680GB of VRAM will be possible
Acho que faria sentido pegar todos esses dados crus que estão publicos hoje e ser feito upload em uma base P2P, provavel que se esse projeto vingar vao derrubar repos, servers, e processar geral, mas se guardar as bases em P2P fica mais safe
Também da para fazer análise de indicativo de corrupcao de licitacoes, eu ja fiz um projeto desse e funcionou mas preferi parar devido ao risco.
Fiz o embedding de todad as licitacoes dos ultimos 3 anos de todos os orgaos do portal da transparencia, e comecei a comparar osvalores