DeepSeek V4.1 Flash 552B (16B active)
on M5 Ultra 512GB · mlx-serve
Summary
User benchmarks DeepSeek-V4.1-Flash on a single M5 Ultra 512GB Mac Studio, averaging 64.1 tok/s generation and 3,101 tok/s prefill over 512 to 256k contexts.
Setup is mlx-serve 26.11.2 with an MLX model; the dual-node run used two M5 Ultra 512GB machines over a single Thunderbolt RDMA link with experts split across nodes.
Two nodes averaged 78.8 tok/s generation (+23%) but prefill fell to 2,812 tok/s (-9%) and 256k TTFT rose from 81.4 to 106.0 s. RDMA was unstable with three queue-pair failures, and the user advises against a dual 256GB Ultra setup in favour of a single 512GB machine.
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
- 73.4 t/s · one run
- 40.0 t/s · one run
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 M5 Ultra 512GB
- 25.7 t/s · one run
- 129.2 t/s · one run
Bars are medians on one scale, 0 to 129.2 tokens/s. Press ▶ to watch text arrive at a row's speed.
Everything reported on the M5 Ultra 512GB →Every DeepSeek V4.1 Flash report →Estimate DeepSeek V4.1 Flash on this card →