Report
Sciforium says DeepSeek-V4.1-Flash hits 1,129 output tokens per second at P90 on four AMD GPUs
Sciforium credits customer-tuned drafters and GPU optimizations for speed gains, with DeepSeek’s original weights unchanged.
TLDR
The company reports 1,129 output tokens per second at P90 on four AMD Instinct MI355X GPUs using a long-context agentic workload. It says customer-tuned speculative drafts let DeepSeek-V4.1-Flash accept more tokens per verification pass. Its estimate of roughly nine seconds for 10,000 tokens covers generation—not the whole job: prompt processing and tool execution are excluded.
Combined views
4.1K
3 Sources, first seen ago

Comments