Expired50% confidenceBenchmarkExact time
On 4×RTX 4090 with prefix caching and DFlash2 speculative decoding both enabled, Qwen3-27B's total throughput drops by approximately 37%.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
9/6/2026
First Seen
Valid until: 9/20/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedThe UP host advises that individual users are not recommended to deploy Qwen3.8-27B, and those with less than 32GB of memory should give up entirely, because the Q4 version occupies approximately 36GB of memory.71% similarUnverifiedWhen deploying Qwen3 27B with 16 GB of VRAM, if some weights are offloaded to system memory, the speed drops sharply from tens of T/s to single digits.71% similarUnverifiedQwen3.6 27B supports a context length of 262K, and the 34B 80-layer version maintains the same context length unchanged69% similarVerifiedQwen2.5-Coder-32B running with Q4 quantization on an RTX 4090 can achieve an inference speed of 40–60 tokens per second.67% similarUnverifiedQwen3-235B-A22B adopts a MoE architecture, activating only 22B out of 235B total parameters at a time.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/864435API
curl https://kongchang.com/api/v1/knowledge/claims/864435MCP
get_claim(id=864435)