Unverified50% confidenceBenchmarkExact time
For the comparison setup Qwen 3.6 35B A3B on llama.cpp, with a VRAM-constrained expert layout and MTP/Vision disabled, prompt processing is 630-680 t/s and generation is 27-35 t/s.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/26/2026
First Seen
Valid until: 11/9/2026
Sources
Related Entities
Related Claims
UnverifiedIn this configuration, Qwen3 27B achieves approximately 70 T/s natively in llama.cpp, with a measured peak of 76 T/s.73% similarUnverifiedCalling Qwen3 27B within an Agent framework (Hornis/赫米) can maintain approximately 60 T/s.72% similarUnverifiedEach of Qwen3-27B's 16 full-attention layers has 4 KV heads and a head dimension of 256, occupying 64 KB per token at 16-bit precision.70% similarUnverifiedOne B70 user running vLLM + GPT-Q INT4 quantization with MTP depth 4 achieved 84.6 tokens/s on a 27B model, but this only supports single-stream and uses different quantization, making direct comparison impossible.70% similarUnverifiedThe official Qwen3-8B configuration has 36 layers, 8 KV heads, head dimension 128, and sliding window attention disabled; with 8192 tokens cached, 2 bytes per element, and a single sequence, the KV cache is about 1.125 GiB.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1099108API
curl https://kongchang.com/api/v1/knowledge/claims/1099108MCP
get_claim(id=1099108)