以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

NVFP4

8 atoms · 跨 5 天 · 首见 2026-05-23 · 最近 2026-06-27

三色: 🟦 fact 6 · 🟥 take 2 · stance ▲4/▼0/◆3

来源: X 8

时态: fresh:8

标签: 好数字:5 · 好观点:2 · 好信源:1

2026-06-27 · 性能退化与改进潜力
NVFP4 has shown a slight performance regression but could probably be fixed with proper L/DoRA and might also totally be worth it given your hardware and serving constraints
tags: 好观点 · → daily
value: direction=轻微退化

2026-05-23 · 即将推出500B模型
We’re getting a 500B NVFP4 model soon guys
tags: 好观点 · → daily
value: qty=500B · date=soon

2026-06-08 · VRAM savings vs FP16
~3-4x VRAM savings vs FP16
tags: 好数字 · → daily
value: qty=3-4x
2026-06-08 · memory bandwidth pressure
Lower memory bandwidth pressure
tags: 好数字 · → daily
value: direction=lower
2026-06-08 · perf per watt
Better perf per watt
tags: 好数字 · → daily
value: direction=better
2026-06-08 · local inference cost
Cheaper local inference
tags: 好数字 · → daily
value: direction=cheaper
2026-06-04 · memory requirement (NVFP4)
350 GB for NVFP4 (read that earlier, haven't checked)
tags: 好数字 · → daily
value: qty=350 · unit=GB
2026-06-02 · format extension described in paper
In this paper is presented MixFP4, a mixed-format FP4 extension built on top of NVFP4, designed to extend the standard NVFP4 format and hardware co-design by enabling an INT-style uniform-step format at block granularity.
tags: 好信源 · → daily
value: qty=MixFP4