Unverified50% confidenceFactExact time
InclusionAI's Ling 3.0 Flash adopts a hybrid MoE design, in which the decoder groups five Kimi Delta Attention (KDA) layers with one Multi-head Latent Attention (MLA) layer, and is natively supported in Transformers with the loading class BailingHybridForCausalLM.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
Sources
Related Entities
Related Claims
Unverified盘古2.0 Flash提供512K的超长上下文窗口，采用MLA、DSA和SWA三种注意力机制的组合设计63% similarUnverifiedStep 3.5 Flash采用混合注意力系统（Hybrid Attention），对短内容使用滑动窗口注意力，对长内容切换为完整注意力机制61% similarUnverifiedJagged Flash Attention (JFA) directly processes variable-length (jagged) sequences, avoiding padding with invalid tokens for alignment.59% similarUnverifiedThe MiniMax H3 architecture uses a unified Transformer backbone to simultaneously handle conditional inputs of different modalities such as text, images, and video, where the 'H' in H3 generally refers to the Hybrid Attention mechanism, combining local window attention with global attention.59% similarUnverifiedGLM-5.3-Flash is open-sourced under the MIT license and supports multimodal inputs including images and tokens.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1039063API
curl https://kongchang.com/api/v1/knowledge/claims/1039063MCP
get_claim(id=1039063)