Home
Chat
Deep Research
News
Pricing
New Chat
Sign in
Sign up
Home
Chat
Deep Research
News
Pricing
New Chat
提
问News
Tech, finance and policy — daily must-reads
Home
Tech
Finance
Policy & Regulation
简体
繁體
EN
Tiwen News
Tags: AMD MI100 inference
AMD MI100 inference
Tag "AMD MI100 inference" — all articles
22
2026-11
Open-source Hy4-Flash debuts with 224B total, 12B active parameters; INT4 version fits fully in 128GB VRAM
Open-source model Hy4-Flash uses a 224B-total, 12B-active MoE architecture, with an early tester reporting a single-pass generation took 32 minutes. The int8-vllm project has released the first INT4 version, which fits fully into 128GB of VRAM, and published benchmark data on older AMD cards including the MI100.
Reddit · 1ncehost、X · LiTianleli