Unverified60% confidenceFactExact time
DeepGEMM-Ascend provides specific operations such as MQA Logit (handling attention scores in multi-query attention) and MegaMode (serving mixture-of-experts MoE models).
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
10/3/2026
First Seen
Valid until: 1/1/2027
Sources
Related Entities
Related Claims
VerifiedDeepSeek adopts a Mixture of Experts (MoE) architecture and a Multi-head Latent Attention (MLA) mechanism.78% similarUnverifiedThe six components have distinct roles: DeepGEMM accelerates general matrix operations, DeepEP handles large-scale cross-device communication, TileKernels provides common vector computation and memory access operators, FlashMLA provides sparse attention operators, and DeepSelect performs efficient data selection.73% similarUnverifiedDeepGEMM is a high-performance matrix multiplication library designed by DeepSeek specifically for its MoE models, supporting low-precision formats such as bf16 and fp8.69% similarUnverifiedDeepSeek-V2引入Multi-Head Latent Attention（MLA）机制和DeepSeekMoE架构，其每百万token推理成本约为同期GPT-4级别模型的百分之一69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/970596API
curl https://kongchang.com/api/v1/knowledge/claims/970596MCP
get_claim(id=970596)