Unverified50% confidenceBenchmarkExact time
On a GPU-less Linux machine (8 vCPU), gemma3:1b takes about 6 to 9 seconds per analysis but often classifies all conversations as asking_questions and gives generic replies; gemma3:4b takes about 15 to 24 seconds per analysis, can select a reasonable stage and reuse facts the seller mentioned, and can reply in Chinese.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGPU 推理速度通常是纯 CPU 的 5-20 倍，当模型太大无法放入显存时部分层会卸载到内存由 CPU 处理（混合运行/partial offload）68% similarVerifiedFor the 4-bit quantized version of a 7B parameter model, at least 8GB of memory and a modern multi-core CPU are typically required for smooth operation, with inference speeds of approximately 10–30 tokens per second67% similarUnverifiedGPU solutions use batching technology to combine multiple user requests for computation, which can greatly increase overall throughput and represents a trade-off between latency and throughput.66% similarUnverifiedGPU推理优化技术如投机解码（speculative decoding）、KV缓存复用对于将端到端延迟压缩到200毫秒以内至关重要65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1033312API
curl https://kongchang.com/api/v1/knowledge/claims/1033312MCP
get_claim(id=1033312)