llamaperf

DeepSeek V4.1 Flash 552B (16B active)

on M3 Ultra 512GB · oMLX · 65,536 ctx

Tone: positiveApple hardware
Nov 4, 2026
Throughput
3123.0 tokens/s prompt processing (prefill)Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
System RAM
512 GB

Summary

User reports DeepSeek V4.1 Flash prefill at 3,123 tok/s with a 64K prompt on an M3 Ultra 512GB paired with an M5 Max 128GB, up from 1,060 tok/s on the M3 Ultra alone. The run uses oMLX 0.7.4.dev1 with its experimental Prefill Accelerator: the M5 Max processes the first layers of each prompt chunk while the M3 Ultra holds the whole model, KV cache and decoding, connected over Thunderbolt or 5GbE. Decode speed is unchanged because decoding still runs on the M3 Ultra. With an M5 Max 64GB instead, the same model reaches 1,466 tok/s (+38%). The post also lists Qwen3.8 Flash-Next at 3,666 tok/s, GLM-5.3 Flash at 1,388 tok/s and MiniMax M3.1 at 569 tok/s with the M5 Max 128GB.

Compare this run

Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.

DeepSeek V4.1 Flash 552B · 16B active at 3-bit by hardware

Bars are medians on one scale, 0 to 73.4 tokens/s. Press ▶ to watch text arrive at a row's speed.

Models most run on the M3 Ultra 512GB

Bars are medians on one scale, 0 to 88.1 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.

Everything reported on the M3 Ultra 512GB →Every DeepSeek V4.1 Flash report →Estimate DeepSeek V4.1 Flash on this card →