Unverified50% confidenceFactExact time
The H3 model is trained with resolutions processed as multiples of 32; keeping Multiple at 32 during inference avoids artifacts or performance degradation caused by size mismatches.
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在27B混合架构、4B稠密模型、36B MoE三类模型上以8k上下文测试，Prefill阶段性能无变化，Decode阶段打补丁后快约1.4%72% similarUnverified千问3.6的35B MOE模型每次只激活其中3B的参数65% similarVerifiedHy3 adopts a Mixture of Experts (MoE) architecture with a total parameter count of 295 billion (295B), activating only 21 billion (21B) parameters per inference.65% similarUnverified研究表明当模型参数规模达到3B以上时，1.58-bit模型在多项基准测试中表现可接近甚至匹配同等规模的全精度模型64% similarUnverifiedAfter testing 30 layer combinations, the best pairing was SenseNova L32 mapped to MiniMax H3 L49, using a 4096→128→5120 projector that achieved a cosine similarity of 0.9042 on the validation set.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/947453API
curl https://kongchang.com/api/v1/knowledge/claims/947453MCP
get_claim(id=947453)