mikedorian369/deepseek-v4-pro-0813-showcase

DeepSeek-V4-Pro-0813 🧠

A showcase Space for deepseek-ai/DeepSeek-V4-Pro-0813, the official 1.65-trillion-parameter mixture-of-experts model with a 1M-token context and enhanced agentic capabilities (built on DeepSeek-V4-Pro with a DSpark speculative-decoding module).

What this Space is

DeepSeek-V4-Pro-0813 is far too large to self-host on a Space (its safetensors metadata reports ~1.6 TB of weights). This Space therefore showcases the model with verified data from its repo — architecture, benchmark table, paper, license, and official usage snippet — and links to real hosted inference via DeepSeek Chat.

Highlights:

• —Architecture cards: total params, hidden size, layers, experts/token, context length, vocab.
• —Benchmark table (HLE, Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon, DSBench…).
• —Verified usage snippet with the model card's recommended sampling parameters.
• —Note explaining the hosting limit and how to get a live chat demo.

Why not a self-hosted demo?

The full model needs multiple GPUs with expert parallelism (~1.6 TB of weights vs. 141 GB on the largest Space GPU). A live chat here would require a DeepSeek API key calling their hosted endpoint (key kept server-side) — available on request.

Links

• —Model card: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
• —Paper: https://arxiv.org/abs/2606.19348
• —Chat: https://chat.deepseek.com/

Built with the Hugging Face Spaces Static SDK.