DeepSeek — MoE with Multi-Head Latent Attention

DeepSeek's MoE family uses Multi-Head Latent Attention (MLA). The full V2/V3/R1 family runs through GPU HIP; distilled variants (Llama- and Qwen-based) add NPU + Vulkan coverage.

Models

Model	Params	1BP Size	Backend(s)	Perf
DeepSeek-V2 / V3 / R1	up to 671B	—	GPU HIP	20 tok/s
DeepSeek-R1-Distill-Llama-8B	8B	4.1 GB	GGML-Vulkan / ZINC / NPU / HIP	44 tok/s
DeepSeek-R1-Distill-Qwen-7B	7B	3.8 GB	ZINC / NPU / HIP	—

Notes

NPU: DeepSeek-R1-Distill-Llama-8B (Llama arch — build_llama), DeepSeek-R1-0528-Qwen3-8B (Qwen3 arch — build_qwen3_8b).
GPU HIP: GGUF validated across the family. Vulkan / CPU: functional.
The distills inherit their base-family tooling — see Llama and Qwen.