DeepSeek V4.1 Flash Runs on One 5090
An independent llama.cpp fork runs DeepSeek-V4.1-Flash on one RTX 5090 paired with 125.7GB of RAM and disk offload. The 552B MoE model reaches 5.12 tokens per second on cold content and 21.27 tokens per second with resident weights.
- –DeepSeek’s 8B prefill and 16B decode activation, plus a KV cache roughly one-quarter the size of the previous model’s, make the architecture unusually efficient for its scale.
DISCOVERED
1h ago
2026-09-15
PUBLISHED
1h ago
2026-09-15
RELEVANCE
AUTHOR
yume_arasaki
