Unverified50% confidenceFactExact time
The new method applies to both continuous and combinatorial inputs, with dimensionality ranging from about 10 to 6000+.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/22/2026
First Seen
Valid until: 1/20/2027
Sources
Related Entities
Related Claims
UnverifiedThe project adopts a mixed-precision quantization strategy: 8-bit for dense layers, 4-bit for expert layers, and 8-bit quantization for the KV cache.64% similarUnverifiedFor models with more than 10B parameters, 4-bit quantization can typically reduce memory usage by approximately 50% while keeping capability loss within 5%.63% similarUnverifiedTaking a layer with a 3×3 convolution kernel and 256 input and output channels as an example, a standard convolution requires approximately 590,000 parameters, while depthwise separable convolution requires only approximately 67,000 parameters, a reduction of roughly 8–9 times.62% similarUnverified位置编码外推方案可将模型有效上下文从最初的4K、8K扩展至128K乃至百万级token62% similarUnverifiedEmbedding模型输出的高维向量通常为768至3072维，向量数据库通过HNSW等图算法可将亿级向量检索压缩至毫秒级响应61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1084381API
curl https://kongchang.com/api/v1/knowledge/claims/1084381MCP
get_claim(id=1084381)