DeepSeek V4 Flash 284B (13B active)

on M5 Max 64GB · ds4

Sep 7, 2026

ShareX Reddit LinkedIn Bluesky

Throughput

30.0 t/s gen

Quant

Q2-Q4 mixed imatrix (GGUF)

System RAM

64 GB

Use cases

agenticcoding

Summary

User reports about 30 t/s on an M5 Max with a quantized DeepSeek V4 Flash running on the DS4 engine. The user states this is double the speed of llama.cpp. The user asks for feedback from CUDA and ROCm users.

View source ↗