I think it's BF16 trained then quantized to FP8, but unsure fully - I was also t... | Hacker News

Hacker Newsnew | past | comments | ask | show | jobs | submit

		danielhanchen 4 months ago \| parent \| context \| favorite \| on: Qwen3-Coder: Agentic coding in the world I think it's BF16 trained then quantized to FP8, but unsure fully - I was also trying to find out if they used FP8 for training natively!

jychang 4 months ago [–]

Qwen uses 16bit, Kimi and Deepseek uses FP8.

danielhanchen 4 months ago | [–]

Oh ok cool thanks!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact