Open-weight & owned stack · Quantization efficiency

Developer reports 1,200-token prefill on Strix Halo

Sep 12, 2026

A community developer reported 1,200 tokens per second prefill for Qwen3.8 Flash Next on Strix Halo.

Read the original at reddit.comOpens the publisher's site in a new tab

More in Open-weight & owned stack

Z.ai publishes GLM-5.3 open weightsSep 8 MiniMax H3 highlighted as open-weight multimodal modelSep 7 InternLM releases 397B multimodal Intern-S2 modelSep 13 DeepSeek Model Nearly Matches GPT-6 Astra DesignSep 10 DeepSeek presents V4 FlashSep 8