The official release of DeepSeek-V4-Pro — a 1.65-trillion-parameter mixture-of-experts model with a 1M-token context and enhanced agentic capabilities, built on the DeepSeek-V4-Pro structure with a DSpark speculative-decoding module.
| Benchmark | V4-Pro-0813 | V4-Flash-0731 | V4-Pro (Prev) | GLM-5.2 | Kimi K3 | Opus-4.8 |
|---|---|---|---|---|---|---|
| HLE (w/ tools) | 60.0 | 51.5 | 48.2 | 54.7 | 56.0 | 57.9 |
| Terminal Bench 2.1 | 87.9 | 82.7 | 72.1 | 81.0 | 88.3 | 85.0 |
| NL2Repo | 61.5 | 54.2 | 38.5 | 48.9 | — | 69.7 |
| Cybergym | 83.3 | 76.7 | 52.7 | — | 80.0 | 78.3 |
| DeepSWE | 62.7 | 54.4 | 12.8 | 46.2 | 67.5 | 58.0 |
| Toolathlon-Verified | 74.1 | 70.3 | 55.9 | 59.9 | 76.5 | 76.2 |
| DSBench-FullStack | 71.1 | 68.7 | 41.8 | 61.8 | 73.7 | 71.6 |
| DSBench-Hard | 67.2 | 59.6 | 31.1 | 54.5 | 63.0 | 71.7 |
Scores from the official model card. Higher is better. DeepSeek-V4-Pro-0813 leads on most agentic / coding benchmarks, with Opus-4.8 competitive on DSBench-Hard.
The tokenizer and model load via transformers. Weights are large (~1.6 TB on disk),
so local deployment needs multi-GPU expert parallelism.
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained(
"deepseek-ai/DeepSeek-V4-Pro-0813"
)
# model = transformers.AutoModelForCausalLM.from_pretrained(
# "deepseek-ai/DeepSeek-V4-Pro-0813", torch_dtype="auto"
# )
# Recommended sampling (from the model card):
# temperature = 1.0
# top_p = 0.95 (agentic) / 1.0 (otherwise)
# max output = 384K tokens for high / max reasoning effortDeepSeek-V4-Pro-0813 cannot be hosted on a Hugging Face Space. Its safetensors
metadata reports roughly 1.65 trillion parameters — about 1.6 TB of weights
(the bulk in int8 expert matrices). The largest Space GPU (NVIDIA H200) has 141 GB of VRAM, so the
full model would need many GPUs with expert parallelism.
This Space therefore showcases the model with verified data from its repo (architecture, benchmarks, paper, license) and links you to the real thing:
Want a live chat demo here? If you have a DeepSeek API key (platform.deepseek.com), I can wire a chat UI that calls the API with the key kept server-side. Just say the word.