返回
原文
🚨 AI News | TestingCatalog@testingcatalog
48AI 编辑部评分，满分 100
2026-08-04 05:30· 19分钟前
跳到正文
AI 摘要

Atomic 在 Hugging Face 发布 DeepSeek V4 Flash 0731 的 14 种量化版本，覆盖无损 BF16 到 1-bit。该模型为 284B 参数 MoE，采用量化感知训练，官方检查点以 MXFP4 存储路由专家。其中 AD-IQ2_M 最适合 128GB 硬件，token 匹配率达 83.6%。

Atomic released 14 quants of "DeepSeek V4 Flash 0731" on Huggingface with the best "quality vs size" performance.

DeepSeek-V4-Flash is a 284B parameter mixture-of-experts model. It is quantization-aware-trained； the official checkpoint already stores its routed experts in MXFP4 and everything else in FP8 or BF16.

The AD-IQ2_M version is the best for testing on 128GB hardware！

atomic.chatRun DeepSeek V4 Flash 0731 locally 🐳 We released 14 quants on Hugging Face, from lossless BF16 to 1-bit AD-IQ2_M is the best fit for 128GB hardware. It matches...
DeepSeekHugging Face开源/仓库开源生态
在 X 查看原推导出 Markdown
返回
48导出 Markdown
🚨 AI News | TestingCatalog · @testingcatalog · X·2026-08-04 05:30·19分钟前
在 X 看原推· x.com
AI 摘要

Atomic 在 Hugging Face 发布 DeepSeek V4 Flash 0731 的 14 种量化版本，覆盖无损 BF16 到 1-bit。该模型为 284B 参数 MoE，采用量化感知训练，官方检查点以 MXFP4 存储路由专家。其中 AD-IQ2_M 最适合 128GB 硬件，token 匹配率达 83.6%。

Atomic released 14 quants of "DeepSeek V4 Flash 0731" on Huggingface with the best "quality vs size" performance.

DeepSeek-V4-Flash is a 284B parameter mixture-of-experts model. It is quantization-aware-trained； the official checkpoint already stores its routed experts in MXFP4 and everything else in FP8 or BF16.

The AD-IQ2_M version is the best for testing on 128GB hardware！

atomic.chatRun DeepSeek V4 Flash 0731 locally 🐳 We released 14 quants on Hugging Face, from lossless BF16 to 1-bit AD-IQ2_M is the best fit for 128GB hardware. It matches...
DeepSeekHugging Face开源/仓库开源生态
在 X 查看原推x.com