DeepSeek V4.1 Flash

Playground UNCENSORED · FP8

Model card ↗

Conversation

Start with a little curiosity.

Explore sample answers, or connect a server running the exact dealignai checkpoint to chat live.

Static preview · samples are not model-generatedEnter to send · Shift+Enter for a new line
Connect the real model · endpoint setup

This free Static Space hosts the interface. Live inference requires your own external server running dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8. No endpoint is supplied and no alternative model is silently used.

  1. Provision suitable external GPU hardware. The model card documents about 510 GB of weights and a tested four-H200 setup.
  2. Serve the exact checkpoint using the model card's SGLang instructions, with --reasoning-parser deepseek-v41.
  3. Expose an OpenAI-compatible HTTPS gateway. Enable CORS for this Space's app origin, shown below, and allow POST, Content-Type, and Authorization. Keep credentials out of public files. Live inference consumes your endpoint's compute budget.
  4. Choose Live model endpoint and enter the API base URL, served model name, and your scoped API key. Changing modes or connection settings clears the chat to keep sessions separate.

App origin: . API keys and conversation history are not saved in browser storage. Your endpoint receives the prompts and may retain logs. This demo supports text chat and displays answer text separately from the server's reasoning channel.

For a server-side secret-backed Gradio deployment after upgrading hosting, see app.py and GRADIO_README.md in the Space's Files tab.