jiuguilaohu2026/kimi-k3

🌙 Kimi K3 — live chat demo

Chat with [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3) — Moonshot AI's open-weights, native-multimodal agentic model:

• —2.8T total / 104B active Stable-LatentMoE (896 experts, top-16 routing)
• —1M-token context, text + image understanding in one model
• —Always-on reasoning with a reasoning_effort dial (low / high / max)
• —Preserved thinking history — prior reasoning is carried into follow-up turns automatically (this demo does it for you)

How this Space works

A 2.8T-parameter model fits on no Space GPU, so there is no backend here at all: this is a static single-page app. Your browser calls [HF Inference Providers](https://huggingface.co/docs/inference-providers) (Fireworks / Featherless) directly, with streaming (SSE).

Who pays? You do, via your own HF token — paste it in the Setup panel (create one at hf.co/settings/tokens; a read-scoped token works). Free monthly HF credits apply. The token never leaves your browser (stored in localStorage only).

Features

• —💭 Live-streamed reasoning block + final answer (Markdown rendered)
• —🖼️ Image upload for K3's native vision (base64 data URLs, nothing stored)
• —⚙️ Reasoning effort / max tokens / temperature / provider selection
• —🔁 Multi-turn with K3's preserved-thinking contract honored
• —🌑 Zero-build single index.html — fork it freely

Built following the huggingface-spaces skill.