Unverified55% confidenceFactTime unknown
Qwen3.5在32K上下文下比上一代Qwen3 Max快8.6倍，在256K上下文下快19倍
1
Sources
55%
Confidence
Medium-term (~90 days)
Relevance
6/1/2026
First Seen
Valid until: 8/30/2026
Sources
Qwen3.5深度解析：混合注意力架构实现19倍长上下文加速
bilibili雪天鱼
Related Entities
Related Claims
UnverifiedQwen3.6 27B支持262K上下文长度，34B 80层版本保持相同的上下文长度不变76% similarUnverifiedQwen3.7 Max的上下文窗口达到256K token75% similarUnverified实测中Qwen3 27B在未开启NVLink的4张魔改2080Ti上生成速度约为40多token/秒74% similarUnverifiedQwen3.6 MTP-GGUF版本在单GPU上将35B-A3B模型的推理速度推到220 token/s，比原版GGUF快超过1.4倍，且精度零损失74% similarUnverifiedQwen2.5-Coder-32B在RTX 4090上以Q4量化运行，推理速度可达每秒40-60个token73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/18076API
curl https://kongchang.com/api/v1/knowledge/claims/18076MCP
get_claim(id=18076)