inference/research
verifiedhigh · 0.86updated 2026-10-18

deepseek-ai/DeepSeek-V4-Pro-0813 model card

repo

Fields

id
sources/hf-deepseek-v4-pro-0813-card
Publisher
DeepSeek
Kind
repo
Published
2026-08-13
Accessed
2026-10-18
notes
Primary card for the DeepSeek-V4-Pro-0813 record, the post-trained checkpoint that supersedes the V4-Pro preview. HF API createdAt 2026-08-13, lastModified the same day, 154,799 downloads and 896 likes at access. Card states the checkpoint keeps the V4-Pro preview's model structure and was post-trained for agentic tasks: Terminal-Bench 2.1 87.9 and Toolathlon Verified 74.1, both marked as the vendor's own testing in the HF eval-results widget. The card recommends vLLM and SGLang and links the SGLang cookbook, and recommends temperature 1.0 with top_p 0.95 for agentic use. License MIT. Safetensors metadata reports 1,650,497,936,906 parameters, dominated by I8 (1,623,497,637,888) plus F8_E4M3 (23,952,621,568), BF16 (2,954,820,352) and F32 (90,530,058), i.e. the same FP4/FP8 mixed release as the preview.

Notes

Primary card for the DeepSeek-V4-Pro-0813 record, the post-trained checkpoint that supersedes the V4-Pro preview. HF API createdAt 2026-08-13, lastModified the same day, 154,799 downloads and 896 likes at access. Card states the checkpoint keeps the V4-Pro preview's model structure and was post-trained for agentic tasks: Terminal-Bench 2.1 87.9 and Toolathlon Verified 74.1, both marked as the vendor's own testing in the HF eval-results widget. The card recommends vLLM and SGLang and links the SGLang cookbook, and recommends temperature 1.0 with top_p 0.95 for agentic use. License MIT. Safetensors metadata reports 1,650,497,936,906 parameters, dominated by I8 (1,623,497,637,888) plus F8_E4M3 (23,952,621,568), BF16 (2,954,820,352) and F32 (90,530,058), i.e. the same FP4/FP8 mixed release as the preview.

Cross-references

This record references no other records.