Skip to content

Model family

DeepSeek v3 / R1

MoE

Notes

Checkpoints listed here identify models in this family. Check each recipe for inference verification, hardware measurements, and requirements before running it.

Feature matrix

Runtimes
generateserve
Backends
CUDA:supportedMetal:supportedROCm:partially supported
Tensor parallelno
Pipeline parallelyes
Draftersnone
KV modes
fp16
Input modalities
text
Outputtokens
Model types
deepseek_v3kimi_k2joyai_llm_flash

Recipes in this family

Choose hardware to estimate fit.