wuisimo/Kimi-K3-Demo

Kimi K3 Demo

Chat with Moonshot AI's Kimi K3 — a 2.8T-parameter open-weight multimodal agentic model (104B activated, 1M-token context).

• —Text + images: ask questions, attach images, or both.
• —Live reasoning: the model's hidden thinking streams in a collapsible panel (Kimi K3 always thinks before answering).
• —Reasoning effort: switch between low, high, and max.
• —Provider choice: route through the HF router (auto), Together AI, Fireworks AI, or Featherless AI.
• —Multi-turn: the preserved-thinking conversation history (including reasoning_content) is passed back to the model on every turn.

How inference is billed

Sign in with your Hugging Face account (OAuth, inference-api scope) — inference requests are made on your behalf and billed to your account via HF Inference Providers. Alternatively, paste a token with the inference-api scope in Settings.

Technical notes

• —Pure static app (no backend): calls https://router.huggingface.co/v1/chat/completions directly from the browser with SSE streaming.
• —Images are compressed client-side (max 1568 px, JPEG) before being sent as data-URL image_url parts.
• —OAuth flow follows the official client-side OAuth example.