Unverified50% confidenceBenchmarkExact time
FlashInfer MNNVL纯AllReduce使DeepSeek-V4-Flash TP4解码在Blackwell架构小批量下获得最高+6.9%性能提升
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/5/2026
First Seen
Valid until: 12/4/2026
Sources
Related Entities
Related Claims
UnverifiedFlashInfer纯allreduce特性已为DeepSeek-V3/V3.2/V4自动启用，其他模型可通过--enable-flashinfer-pure-allreduce手动开启76% similarUnverifiedDeepSeek V4 Flash是Pro的蒸馏或精简版本，在保留核心编程能力的同时大幅降低了推理开销72% similarVerifiedDeepSeek将实验性多模态模型V4-Flash-Vision-EXP上线API平台，其文本能力与V4 Flash持平，多模态智能体基准测试性能接近Opus 4.8水平70% similarUnverifiedDeepSeek-V4-Flash-Vision-Exp采用MIT许可开放权重，并提供Tensor与Tokenizer参考实现及最小化PyTorch推理代码69% similarUnverifiedDeepSeek在7月的更新日志中将一个内部Harness（称为「极简模式」）归功为V4 Flash编程基准评分的幕后功臣68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/860551API
curl https://kongchang.com/api/v1/knowledge/claims/860551MCP
get_claim(id=860551)