Expired50% confidenceBenchmarkExact time
When VRAM is sufficient and the model is fully resident on the GPU, an RTX 4090 running a Q4-quantized 27B model can typically achieve 15–30 tokens/s; once CPU offloading is triggered, the speed may drop sharply to 1–2 tokens/s or even lower.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/12/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/10/2026(expired)
Sources
Related Claims
UnverifiedOn the RTX 5090, the Qwen3.6 27B model's prefill speed increases from 174 to 253 tokens/s, with a VRAM usage of approximately 18 GB.81% similarUnverifiedFor the RTX 4090 (24GB VRAM), WhichLLM may recommend the Q5 quantized version of Qwen3.6 27B, with an estimated inference speed of approximately 27 tokens per second.77% similarUnverifiedRunning the Qwen3.6 27B model on a GPU with 24 GB of VRAM, the generation speed reaches approximately 40 tokens per second without optimization, and 50–60 tokens/s after optimization.77% similarUnverifiedOn the RTX 5090, the Qwen3.6 27B model with Q4 quantization has a decoding speed of approximately 63 tokens/s without MTP, jumping to 105 tokens/s with MTP enabled.76% similarUnverifiedOn the RTX 5090, the Qwen3.6 35B-A3B achieves an inference speed of 220 tokens/s with approximately 23 GB of VRAM usage.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/489738API
curl https://kongchang.com/api/v1/knowledge/claims/489738MCP
get_claim(id=489738)