Unverified50% confidenceBenchmarkExact time
Haiku 5.5在不同供应商处可跑到100到200 TPS，创作者用Claude Code测试稳定在约180 TPS
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/9/2026
First Seen
Valid until: 1/7/2027
Sources
Related Entities
Related Claims
UnverifiedClaude Haiku 4.5拥有20万token的上下文窗口和6.4万token的最大输出长度70% 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 TPS62% similarUnverified支持1080P@60Hz传输的Miracast投屏器实际有效码率约需20Mbps以上稳定带宽，2.4GHz频段因干扰难以持续满足，选择5GHz频段设备可显著减少卡顿61% similarVerifiedQwen2.5-Coder-32B在RTX 4090上以Q4量化运行，推理速度可达每秒40-60个token61% similarUnverified在AMD RX 7900XTX 24GB显卡上运行Qwen3 27B级别模型，输出速度最高可达52~53 TPS，平均稳定在47 TPS左右60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990912API
curl https://kongchang.com/api/v1/knowledge/claims/990912MCP
get_claim(id=990912)