Unverified50% confidenceBenchmarkExact time
3-bit Qwen3.8-27B decodes at 153 tok/s on the tester's engine (compared with 132 tok/s), but each task still takes 18 minutes instead of 13 minutes.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/2/2026
First Seen
Valid until: 11/16/2026
Sources
Related Entities
Related Claims
UnverifiedWhen running IQ3_XXS quantized models with llama.cpp, the output speed is only 15 tok/s with prompt processing at 100–120 tok/s, while a self-developed engine brings the same quantization up to approximately 65 tok/s output and 430 tok/s prompt processing, representing close to a 4x performance improvement.77% similarUnverifiedWhen Qwen3 27B used Three.js to generate a 3D racing game with the context set to 65,000, it took 2 hours and 15 minutes, averaging only 4.78 tokens per second.76% similarUnverifiedOverlapped checkpoint loading for Qwen3-32B on H100 achieves a 2.38× speedup compared to the default plain loading method (35.6 seconds vs. 84.8 seconds).74% similarUnverifiedAt approximately 4K input, TTFT is 23.6 seconds with decoding at 46.6 tok/s; at 8K, 47 seconds and 41.4 tok/s; at 16K, 1 minute 30 seconds and 30.8 tok/s; at 30K, 3 minutes 15 seconds and 31.9 tok/s.73% similarVerifiedOn a T4 GPU, Laya takes 33 milliseconds per single question, and batch processing can compress that to 7 milliseconds72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1127043API
curl https://kongchang.com/api/v1/knowledge/claims/1127043MCP
get_claim(id=1127043)