DeepSeek releases DeepEP V2, TileKernels

This is infrastructure news masquerading as a repo drop: DeepSeek is exposing more of the low-level machinery behind efficient frontier-scale MoE systems.

• –DeepEP V2 claims up to 1.3x peak performance versus V1 while cutting SM usage by up to 4x, which matters when communication and compute need to overlap cleanly.
• –Switching from NVSHMEM to a lighter NCCL Gin backend should make integration less exotic for teams already built around NCCL communicators.
• –TileKernels is early and lightly documented, but its MoE, FP8/FP4 quantization, Engram, and mHC kernels offer a useful glimpse into DeepSeek’s internal training and inference priorities.
• –The SM100/Blackwell target is notable because it signals where serious MoE optimization work is moving, even if most developers cannot run this stack locally yet.
• –For AI infra teams, this is less a plug-and-play release than a benchmarkable reference point for how DeepSeek thinks about model-system co-design.