Unverified50% confidenceBenchmarkExact time
运行Qwen 3（27B）无优化基线token解码速度：RTX 5090约75、RTX 4090约46、RTX 3090约40、M5 Max 31.4、DGX Spark 12-17、Ryzen AI Max 11.5 tokens/s
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/4/2026
First Seen
Valid until: 1/2/2027
Sources
Related Entities
Related Claims
VerifiedQwen2.5-Coder-32B在RTX 4090上以Q4量化运行，推理速度可达每秒40-60个token78% similarUnverifiedQwen 27B的Q4_K_M/GGUF版本在同一台机器上解码速度约为10.7 tok/s，约为MLX版本的1/3.878% similarVerified在RTX 5070（12GB显存）配置下，本地Qwen3 9B模型的输出速度达到约70-76 token/秒77% similarVerified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76% similarUnverified在一台128G算力舱设备上，Qwen3.8 Flash的解码速度常态达241 TPS，峰值达274 TPS75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/973165API
curl https://kongchang.com/api/v1/knowledge/claims/973165MCP
get_claim(id=973165)