OpenAI-compatible inference · now in invite-only beta

The open-weights models,
one API away.

Serve Llama, Gemma, Qwen, DeepSeek and Kimi through a single OpenAI-compatible endpoint. Serverless GPUs, token-accurate billing, keys you control.

chat.py
# pip install openai
from openai import OpenAI

client = OpenAI(
  base_url="https://api.aicloud.dev/v1",
  api_key="sk-ac-live-…",
)

stream = client.chat.completions.create(
  model="llama-3.1-8b-instruct",
  messages=[{"role": "user", "content": "Ship it."}],
  stream=True,
)
Model catalog

Five families. One endpoint.

From an 8B you can stream in real time to 8×-GPU giants that scale from zero — pick a slug and go.

Llama

3.1 8B · 3.3 70B

Gemma

3 27B IT

Qwen

3 32B · Embeddings

DeepSeek

V3 · scale-from-zero

Kimi

K2 · 8× GPU giant

Embeddings

Qwen3-Embedding

A control plane built for developers.

Everything you need to run production inference — and nothing you don’t.

Drop-in OpenAI API

Point the OpenAI SDK at one base URL. Chat, streaming and embeddings — no code rewrite.

Serverless GPUs

Autoscaling replicas on datacenter-grade hardware. Giants cold-start from zero on demand.

Token-accurate metering

Every prompt and completion token counted, with per-model and daily breakdowns.

Org-scoped keys

Issue and revoke keys per organization, with per-key rate limits. Revocation is instant.

Invite-gated & private

Sessions handled by Ory Kratos. Prompts are never persisted on GPU workers.

Streaming-first

Time-to-first-token under 2s on small models. SSE relayed verbatim, no buffering.

< 2s
Time-to-first-token
99%
30-day request success
6
Launch models
/v1
OpenAI-compatible

Have an invite code?

Create your organization, mint a key, and make your first call in under two minutes.