DeepSeek has released the weights for DeepSeek V4.1-Flash on Hugging Face under the deepseek-ai account. The announcement quickly became one of the top posts on r/LocalLLaMA.

The model is a sparse mixture-of-experts design. Community reports cite a backbone in the range of 485B to 552B total parameters, with only 8B parameters active per token. The architecture uses 1 shared expert and 384 routed experts per MoE layer, activating 6 routed experts per token.

Additional components include Single-Pass mHC, a revised residual-stream mixing with a Mega-mHC kernel, an Engram conditional memory of 196B parameters accessed sparsely via token-based lookup, and DSpark speculative decoding with semi-autoregressive draft generation. The context window is reported at one million tokens.

The "Flash" name refers to cheap, fast inference: only 8B active parameters translate to low API serving costs. Local hosting is another matter. One commenter with a 256GB Xeon server estimates needing 384GB of RAM. The model does not fit dual DGX Spark systems either. Users are already discussing NVFP4 quantization while keeping the 196B Engram component in memory.

By the sqrt(P x A) metric cited in the thread, competence should roughly match a similarly trained 62B dense model. Some commenters say they would prefer a dense variant at that size. Early users nevertheless report fast, high-quality answers even on a Mac Studio.