Unverified50% confidenceBenchmarkExact time
In concurrent request scenarios sharing the same system instructions, RadixAttention can reduce prefill-stage computation by more than 80%.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedContinuous batching immediately removes completed sequences and inserts new requests after each decoding step, raising GPU utilization from 30–50% in traditional approaches to over 80%.66% similarUnverified通过精简上下文、优化System Prompt设计、采用少样本示例等Prompt工程技巧，可将相同质量输出的Token消耗降低30%-60%63% similarUnverified据实测数据，约50-70%的生产请求可被本地模型以可接受质量完成，混合路由通常可将总推理费用降低40-60%，并将平均响应延迟压缩到纯API方案三分之一以下63% similarUnverifiedIndustry estimates suggest that implicit tokens (system prompts, RAG retrieval context, function descriptions, and conversation history) typically account for 60%–80% of total consumption per request.62% similarUnverified少样本方式在格式化输出、分类任务和规则遵循方面的幻觉率可降低40%-70%62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/893349API
curl https://kongchang.com/api/v1/knowledge/claims/893349MCP
get_claim(id=893349)