Device-specific formats · 4K context
DeepSeek-V4.1-Flash MoE model with 552B total parameters and 16B active per token · 264.5GB Q2_K model file · working memory unconfirmed
An MoE model with a 552B backbone and separate Engram tables. Published runs on Spark, RTX 5090, and high-memory Macs are separated by hardware and format, including experimental paths that require SSD streaming and dedicated runtimes.
- Q2_K model file
- 264.5GB · memory-mapped paging
- total parameters
- 552B
- active parameter
- 16B
- native context
- 1024K
Representative equipment to run this model
Decode
2.8 tok/s · Published CUDA-cMoE reference
First token
2155.8 s · Not a direct TTFT measurement
available 480GB · 819GB/s
oMLX 0.7.0.dev2 · Lightning MTP on · Engram resident in RAM
Decode
23.1 ~ 41.0 tok/s
First token
7.1 to 12.6 seconds
DeepSeek-V4.1-Flash Local Serving Run Recipe
Selecting a device changes the quantization, safety context, GPU offload, Flash Attention, KV cache, and batch start values.
If you want to choose equipment again
You can first narrow down the model configurations based on budget, noise level, operating system and whether it's acceptable used.