Unverified50% confidenceFactExact time
Qwen3.8-Flash-Next uses a hybrid attention mechanism combining GDN and QSA, with the MoE layers using 512 experts and Top-10 routing.
1
Sources
50%
Confidence
Long-term
Relevance
9/6/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedQwen3.8-Flash-Next adopts a hybrid attention mechanism of GDN plus QSA, with a total of 125B parameters plus 51B N-gram embeddings, activating only approximately 6B parameters per token.80% similarUnverifiedGQA（分组查询注意力）等技术可以有效压缩KV Cache，Qwen3系列也采用了类似优化69% similarUnverifiedGLM-5.3-Flash adopts a hybrid architecture of sparse attention and linear attention, and runs entirely on a domestically produced AI chip cluster.68% similarVerifiedThe Qwen 3.5 architecture uses an improved Transformer variant that incorporates technologies such as Grouped Query Attention (GQA) and Rotary Position Embedding (RoPE).67% similarUnverifiedLLaMA-2 70B采用80层、64个注意力头的架构，并使用GQA（8个KV头共享64个Q头）67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/863612API
curl https://kongchang.com/api/v1/knowledge/claims/863612MCP
get_claim(id=863612)