Gully Labs' own LLM

Open-weight AI, served by Gully

GullyAI runs Qwen3.6-27B (Apache-2.0 open weights) on vLLM, behind an OpenAI-compatible API with a key, a rate limit and a token meter for every product. One model for Gully Street, the Telegram bots, Gully Security, Gully Studio and HolderMaps.

Base URL
https://ai.gullylabs.xyz/v1
Model
gully-ai
Gateway status
GET ai.gullylabs.xyz/healthz
Checking…Asking the gateway now.

Each dot is one health check made by this page, every 30 seconds. Nothing here is simulated.

The GullyAI robot mascot typing on a laptop
What you get

An API you already know

If your code talks to OpenAI, it talks to GullyAI. Change the base URL and the key, keep everything else.

OpenAI-compatible

Point the OpenAI SDK at ai.gullylabs.xyz/v1 and send model gully-ai. Chat completions, streaming, /v1/models and /v1/usage.

Per-product API keys

Every product gets its own gai_ key. Only a SHA-256 hash of it is stored, and one key can be revoked without touching the others.

Rate limits and metering

Requests per minute, tokens per minute and a max-token cap per key. Tokens are metered per key. Prompts and completions are never logged.

Long context

131,072 tokens configured on a 48 GB GPU (FP8 weights). The model itself supports 262K natively on larger cards.

Tool calling

Send tools in the OpenAI format and get tool_calls back, streaming included. Built for agents that call Gully services.

Reasoning

The model thinks before it answers. The thinking comes back in reasoning_content, separate from the answer, and can be turned off for fast, cheap calls.

Quickstart

Three settings, one call

Base URL, your key, model gully-ai. The same request works from curl, the OpenAI SDK for Node or Python, or any OpenAI-compatible client.

curl https://ai.gullylabs.xyz/v1/chat/completions \
  -H "Authorization: Bearer $GULLY_AI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gully-ai",
    "messages": [{"role": "user", "content": "Explain gas in one sentence."}],
    "max_tokens": 300
  }'
Across Gully

One model for every Gully product

Each product calls GullyAI from its own backend with its own key, so limits and usage are per product and nothing leaks into a browser or an app.

Gully Chain · chain 485590

Agents that act on Gully Chain

Gully Chain is supported in every Gully product. GullyAI's part is the agent: a tool-calling loop that reads the chain through Gully RPC, checks tokens with HolderMaps and Gully Security, and pays for paid endpoints with x402 through Gully Pay. Money and chain writes always need a confirm or a hard budget.

The agent is designed, not built yet. It comes after GullyAI has served plain chat for a few weeks.

  • Reads through Gully RPCBalances, calls and receipts on Gully Chain (chain 485590) and every other Gully network, with a Gully RPC key.
  • Holders and scansTop holders, clusters and risk flags of Gully Chain tokens from HolderMaps and Gully Security.
  • Pays with x402Paid Gully endpoints are paid per request in glUSD on Gully Chain, through Gully Pay. A hard budget per run.
Where it stands

Built, tested, waiting for its GPU

Straight answers about what exists today.

  1. Gateway built and testedKeys, limits, metering and streaming, tested end to end against a mock model server.
  2. Model and engine chosenQwen3.6-27B FP8 on vLLM, pinned versions, a one-command deploy.
  3. 04Live for Gully productsKeys go out to Gully Street, the bots, Security, Studio and HolderMaps.
Open to everyone

Build with GullyAI.

No waitlist. Tell us what you are building and we issue your key with its own limits. It works as soon as the model server is online.