---
title: Kimi K3 Console
emoji: 🌙
colorFrom: indigo
colorTo: blue
sdk: static
app_file: index.html
pinned: false
short_description: Chat with Kimi K3 — 2.8T MoE, vision, 1M context
---

# 🌙 Kimi K3 Console

A browser-only chat client for [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3).

K3 is a 2.8T-parameter MoE (104B active per token) with native vision and a
1M-token context window. Serving those weights takes a multi-node cluster, so
this Space doesn't try — it calls a hosted API straight from your browser.
No backend, no GPU, and your API key never leaves your machine.

## Getting a key

Pick either provider in the settings panel. Both are OpenAI-compatible and both
send CORS headers, which is what makes the browser-direct approach work.

| Provider | Endpoint | Model id | Cost |
|---|---|---|---|
| Moonshot (official) | `https://api.moonshot.ai/v1` | `kimi-k3` | $1 minimum top-up |
| OpenRouter | `https://openrouter.ai/api/v1` | `moonshotai/kimi-k3` | ~$3 / M in, $15 / M out |

Neither has a free tier. Keys are held in memory unless you tick
**Remember on this device**, which stores it in `localStorage` — your browser
only, never transmitted anywhere except to the provider you chose.

## Two K3-specific details this client handles

**Preserved thinking history.** K3 always reasons, and it was trained expecting
the complete assistant message — `reasoning_content` included — to be passed
back verbatim on the next turn. A wrapper that keeps only `content` quietly
degrades multi-turn quality. This client keeps the wire history separate from
what it renders, so the trace survives the round trip.

To watch it work, send *"Tell me three random numbers"*, then ask which ones it
held back. The follow-up only resolves if the model can still see its own
scratchpad.

**Base64 vision input.** K3 rejects public image URLs, so attachments are
inlined as `data:` URLs. Paste or drop images straight into the composer.

Thinking effort is exposed as `low` / `high` / `max`. There is no "off".

## The ribbon

The band under the header is 896 ticks with exactly 16 lit — K3's real routing
sparsity — reshuffling while tokens stream. It's the generating indicator.

## Running locally

It's one file with no build step and no dependencies:

```bash
python3 -m http.server 8000
```

Then open <http://localhost:8000>.