🤖 UVR LLM Dashboard

Enter password to access

Server runs 24/7. No wake-up needed — requests route to vLLM directly.

🌐 Web Chat (Open WebUI)

https://ai.uvrlab.org

• Signup password: kaistuvr
• Register with your email and the signup password

📡 API Endpoints

• Primary gateway: https://aispark.uvrlab.org/v1 — OpenAI-compatible. Serves DeepSeek-V4-Flash from the local 2× GB10 Spark cluster and relays the Qwen / Gemma models.
• Alternate host: https://ai.uvrlab.org — Open WebUI web chat (above) and the GPU host that physically serves the Qwen / Gemma models the gateway relays.
• API Key: your personal token (see wiki link above) — sent as Authorization: Bearer <token> or x-api-key: <token>.
• Format: OpenAI-compatible — /v1/chat/completions, /v1/models; streaming & tool-calling supported.

🤖 Available Models

• deepseek-v4-flash — DeepSeek-V4-Flash 284B/13B MoE, TP=2 (2× GB10 Spark), ~40 tok/s, 🔧 tools, 💭 reasoning, MTP, 1M context ✅ (default · self-hosted)
• Qwen3.6-27B-MTP — Qwen3.6 27B MTP, reasoning, speculative decode (hosted on ai.uvrlab.org)
• gemma-4-12B — Gemma 4 12B, fast (hosted on ai.uvrlab.org)

🖥️ Cursor IDE Setup

1. Open Cursor Settings → Models → OpenAI API Key section
2. Set Base URL to: https://aispark.uvrlab.org/v1
3. Set API Key to your personal token
4. Set Model name to: deepseek-v4-flash
5. Click "Verify" to test the connection

Supports both chat and agent mode (tool calling enabled).

💻 cURL Example

curl https://aispark.uvrlab.org/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 1024,
    "stream": true
  }'

🐍 Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI(
    base_url="https://aispark.uvrlab.org/v1",
    api_key="YOUR_API_KEY"
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1024,
    stream=True
)
for chunk in response:
    print(chunk.choices[0].delta.content or "", end="")

🧩 opencode config

Add to ~/.config/opencode/opencode.json. Two providers so each model uses its best endpoint (DeepSeek → Spark gateway; Qwen/Gemma → ai.uvrlab.org direct). Replace YOUR_API_KEY with your personal token.

{
  "provider": {
    "uvrlab-spark": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "UVR Lab — Spark",
      "options": {
        "baseURL": "https://aispark.uvrlab.org/v1",
        "apiKey": "YOUR_API_KEY"
      },
      "models": {
        "deepseek-v4-flash": {
          "name": "DeepSeek V4 Flash",
          "attachment": false,
          "tool_call": true,
          "limit": { "context": 1048576, "output": 131072 },
          "cost": { "input": 0, "output": 0, "cache_read": 0 }
        }
      }
    },
    "uvrlab-ai": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "UVR Lab — AI",
      "options": {
        "baseURL": "https://ai.uvrlab.org/v1",
        "apiKey": "YOUR_API_KEY"
      },
      "models": {
        "Qwen3.6-27B-MTP": {
          "name": "Qwen3.6 27B MTP",
          "attachment": false,
          "tool_call": true,
          "limit": { "context": 64000, "output": 64000 },
          "cost": { "input": 0, "output": 0, "cache_read": 0 }
        },
        "gemma-4-12B": {
          "name": "Gemma 4 12B",
          "attachment": false,
          "tool_call": false,
          "limit": { "context": 256000, "output": 256000 },
          "cost": { "input": 0, "output": 0, "cache_read": 0 }
        }
      }
    }
  }
}