Unverified50% confidenceOpinionExact time
In the past, optimization of on-device large models focused mostly on quantization and pruning, but system-level scheduling and prefetching can likewise unlock enormous headroom.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
iPhone Runs 35B MoE at 20.5 tok/s: The Scheduling Magic That Skips No Experts
redditr/LocalLLaMA10/21/2026
Related Entities
Related Claims
UnverifiedVRAM capacity often determines what size model can be loaded more than raw compute performance does.71% similarUnverifiedWhen a model contains a large number of scale-type parameters (such as covariance or rate parameters), especially in high-dimensional settings, it is advisable to consider optimization strategies that more closely follow the Fisher–Rao geodesic.71% similarUnverified通过提供更多具体案例（few-shot prompting）约束模型输出空间，是纠正AI过度设计的有效手段70% similarUnverifiedIn RAG, parameters such as chunk size, overlap length, embedding model selection, similarity threshold, and the value of K all significantly affect the final outcome.70% similarUnverified叠加假说（Superposition Hypothesis）认为模型在有限维度中压缩存储远多于维度数量的特征，并通过近似正交的方向加以区分70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1081665API
curl https://kongchang.com/api/v1/knowledge/claims/1081665MCP
get_claim(id=1081665)