Expired90% confidenceFactExact time
本次补丁将nvidia-cutlass-dsl默认设置为cu13，修复了Qwen 3.5模型在FlashAttention-4中的崩溃问题
1
Sources
90%
Confidence
Medium-term (~90 days)
Relevance
5/27/2026
First Seen
Valid until: 8/25/2026(expired)
Sources
SGLang v0.5.12.post1 Released: DeepSeek V4 Stability Fixes and Blackwell Adaptation
twitterlmsysorg5/27/2026
Related Entities
Related Claims
UnverifiedQwen3 27B是原生多模态模型，支持视觉输入，弥补了DeepSeek V-Flash无法处理图像的短板66% similarUnverifiedSeries such as Qwen2.5 and Gemma3 have already released 1B–3B quantized versions optimized specifically for edge scenarios.63% similarUnverifiedSGLang currently cannot run the Qwen3 27B NVFP4 MTP-XS model, whereas vLLM already supports it.61% similarUnverifiedThe OPUS 4.6 distilled version comes from the Jacker community and performed well in the Qwen 3.5 era, but the 3.6 V1 preview test results were unsatisfactory, and its Q6 precision also falls short of NVFP4.61% similarUnverifiedQwen3 Max 更新后工具调用稳定性大幅提升，跳过文件创建的老问题已基本消除61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/565API
curl https://kongchang.com/api/v1/knowledge/claims/565MCP
get_claim(id=565)