DeepSeek-V4.1-Flash hits 1,129 tokens/sec on AMD
Sciforium engineers achieved 1,129 tokens/sec with DeepSeek-V4.1-Flash using four AMD MI355X GPUs and a secondary speculative decoding model.
Speculative decoding continues to be the most practical path to massive inference speedups on hardware like AMD's. Pushing over 1,000 tokens/sec on four GPUs suggests massive efficiency gains for real-time applications, and this approach highlights how model architecture and inference optimization can outpace pure hardware scaling.
DISCOVERED
2h ago
2026-11-17
PUBLISHED
3h ago
2026-11-17
RELEVANCE
AUTHOR
techNmak