[KongchangAI]
Unverified50% confidenceFactExact time

The BF16 weights of the 27B model are approximately 54GB (27B × 2 bytes).

1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen

Sources

Can a 27B FP8 Model Fit on a 24GB GPU? Breaking Down the Real VRAM Requirements

Related Entities

BF16technology

Related Claims

UnverifiedM3原始权重存储下限：BF16约854GB，8-bit约428GB，4-bit约214GB80% similarVerified32B参数规模的模型在量化后通常可以压缩到约16-20GB显存占用，一张NVIDIA RTX 4090（24GB显存）就能运行79% similarVerified一个原本需要60GB显存的30B参数模型，经过量化后能在16GB甚至8GB显存的消费级显卡上运行79% similar

Cite This Claim

Stable URI
https://kongchang.com/claim/828258
API
curl https://kongchang.com/api/v1/knowledge/claims/828258
MCP
get_claim(id=828258)