Expired50% confidenceReleaseExact time
This week, an unreviewed 27B FP8 weight version model appeared on HuggingFace.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
8/31/2026
First Seen
Valid until: 9/14/2026(expired)
Sources
Related Entities
Related Claims
Unverified此次工作采用FP8 rollout搭配BF16 actor的混合精度方案54% similarVerifiedFP8 is an 8-bit floating-point format natively supported by NVIDIA's Hopper architecture (H100), which preserves better dynamic range compared to INT8 quantization.54% similarUnverified许多HuggingFace模型仓库已经提供了ONNX版本，特别是针对CPU和移动端优化的量化模型53% similarUnverifiedFP8由NVIDIA在Hopper架构（H100）中引入硬件原生支持，随后在Ada Lovelace（RTX 40系）和Blackwell（RTX 50系）架构中持续完善52% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/832462API
curl https://kongchang.com/api/v1/knowledge/claims/832462MCP
get_claim(id=832462)