Live at constantinople.cloud — OpenAI-compatible API powered by decentralized GPU miners.

Hidden state verification protocol for trustless AI inference on Bittensor.

Use the API with any OpenAI-compatible client:

curl https://api.constantinople.cloud/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer cst-YOUR_API_KEY' \
  -d '{
    "model": "Qwen/Qwen2.5-7B-Instruct",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 256,
    "stream": true
  }'

from openai import OpenAI

client = OpenAI(
    base_url="https://api.constantinople.cloud/v1",
    api_key="cst-YOUR_API_KEY",
)

for chunk in client.chat.completions.create(
    model="Qwen/Qwen2.5-7B-Instruct",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
):
    print(chunk.choices[0].delta.content or "", end="")

Get an API key at constantinople.cloud/api.

Client → POST /v1/chat/completions (OpenAI-compatible)
       → Proxy Gateway (stateless forwarder, routes + logs)
       → Miner (serves inference + returns hidden state commitments)
       → Audit Validator (async challenges, cosine similarity check)
       → Scoring Engine (speed × verification × consistency)
       → On-chain weight setting via commit-reveal

Miners prove they're actually running the LLM by returning hidden state vectors that match the validator's reference model (cosine similarity > 0.70 to pass, > 0.99 for full credit).

The validator runs as two split services:

• Proxy Gateway (proxy_gateway.py): Stateless request forwarder with metagraph discovery, inline commitment injection, R2 audit logging, and wall-clock TPS validation.
• Audit Validator (audit_validator.py): Async auditor that reads R2 records, sends deferred hidden state challenges (30-180s delay), runs reference forward passes, and sets on-chain weights.

# Requires: wallet registered on subnet 97
docker compose up -d
# Watchtower auto-pulls new images every 5 minutes

Miners serve LLM inference with hidden state extraction for verification. The current required model is Qwen/Qwen2.5-7B-Instruct.

# vLLM miner (recommended, high throughput)
python vllm_miner.py --port 8091 --model Qwen/Qwen2.5-7B-Instruct --gpu-memory-utilization 0.70 --hf-device cpu

# Multi-GPU
python vllm_miner.py --port 8091 --model Qwen/Qwen2.5-7B-Instruct --tensor-parallel-size 2

# HuggingFace miner (simpler, lower throughput)
python real_miner.py --port 8091 --model Qwen/Qwen2.5-7B-Instruct

See MINER_GUIDE.md for detailed setup instructions, scoring breakdown, and troubleshooting.

See THREAT_MODEL.md for attack analysis and defense details (29+ vectors analyzed).

MIT