DeepSeek-V4 series model card (V4-Pro and V4-Flash)

Fields

id

sources/model-cn-deepseek-v4-card

URL

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash

Publisher

DeepSeek

Kind

release-notes

Published

2026-04-22

Accessed

2026-10-03

notes

Card table: DeepSeek-V4-Pro 1.6T total / 49B activated; DeepSeek-V4-Flash 284B total / 13B activated; both 1M context, trained on 32T+ tokens. Also gives the lineage row 671B/37B for V3. V4-Flash ships with expert_dtype fp4 and a quantization_config of fp8 e4m3 with weight_block_size 128x128, so the released weights are pre-quantized rather than bf16. Both V4-Pro and V4-Flash are MIT. Note the V4 config.json has NO kv_lora_rank - the latent width is not stated there, which is why the MLA KV derivation for V4 is left null.

Notes

Card table: DeepSeek-V4-Pro 1.6T total / 49B activated; DeepSeek-V4-Flash 284B total / 13B activated; both 1M context, trained on 32T+ tokens. Also gives the lineage row 671B/37B for V3. V4-Flash ships with expert_dtype fp4 and a quantization_config of fp8 e4m3 with weight_block_size 128x128, so the released weights are pre-quantized rather than bf16. Both V4-Pro and V4-Flash are MIT. Note the V4 config.json has NO kv_lora_rank - the latent width is not stated there, which is why the MLA KV derivation for V4 is left null.

Cross-references

This record references no other records.