Unverified90% confidenceFactTime unknown

GPT-OSS原生支持FP4混合精度，无需额外量化即可大幅压缩显存占用

1

Sources

90%

Confidence

Long-term

Relevance

6/1/2026

First Seen

Sources

OpenAI开源GPT-OSS：16G显存跑O4级模型，部署教程全解析

bilibili九天Hector

Related Entities

GPT-OSSmodel FP4concept

Related Claims

UnverifiedGPT-2（355M）在混合精度（FP16）下微调显存占用约2-4GB，单卡T4即可完成全流程82% similar UnverifiedGPT-4o 在保持 128K 上下文能力的同时实现了更快的推理速度和更低的成本76% similar UnverifiedGPT-4o-mini推理速度快、价格低廉，适合日常轻量编码和代码补全76% similar UnverifiedGPT-OSS 20B版本推理时仅激活3.6B参数，性能接近O3-mini73% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/23207

API

curl https://kongchang.com/api/v1/knowledge/claims/23207

MCP

get_claim(id=23207)