DeepSeek published open weights for both V4 models on Hugging Face under an MIT license, continuing the open-weights pattern from V3 and R1.

The specs that matter for self-hosting: V4 Pro is a 1.6T-parameter MoE with 49B active parameters; V4 Flash is 284B total with 13B active. Both list a 1M-token context window and a 384K max output.

At 284B total parameters, Flash is 'small' only by comparison — full-precision weights need roughly 170 GB of memory, and even aggressive quantization keeps you above 30 GB of VRAM. See our local setup guide for the tier-by-tier breakdown.

Sources and publication record

Source material is linked for readers who want to verify the underlying announcement or documentation.