OpenAI-compatible inference

One API. Local models, served by real hardware.

Point your OpenAI client at gigatoken and route requests across a pool of provider-run devices. Pick a model for what you're building, get a key, and ship.

Get API key Read the docs

Pick a model for what you're building

Each family has Fast, Standard, and Smartest picks ordered by latency and quality. Pass the alias in the model field; each cell includes the model's minimum VRAM hint for providers.

Family	Fast Lowest latency	Standard Balanced quality	Smartest Highest quality
Prose Everyday chat, assistants, and character/story roleplay Open WebUI, SillyTavern, chat front-ends	Gemma 4 26B A4B Q4 `google/gemma-4-26b-a4b-it-q4` 26B / 4B active / Q4 Min VRAM: 19 GiB	Gemma 4 26B A4B `google/gemma-4-26b-a4b-it` 26B / 4B active Min VRAM: 28 GiB	Gemma 4 31B `google/gemma-4-31b-it` 31B params Min VRAM: 36 GiB
Code & agents Code generation and editing, and tool-calling autonomous agents opencode, Kilo Code, OpenClaw	Qwen3.6 35B A3B Q4 `qwen/qwen3.6-35b-a3b-q4` 35B / 3B active / Q4 Min VRAM: 23 GiB	Qwen3.6 35B A3B `qwen/qwen3.6-35b-a3b` 35B / 3B active Min VRAM: 37 GiB	Qwen3.6 27B `qwen/qwen3.6-27b` 27B params Min VRAM: 31 GiB

Prose

Everyday chat, assistants, and character/story roleplay

Open WebUI, SillyTavern, chat front-ends

FastLowest latency

Gemma 4 26B A4B Q4

google/gemma-4-26b-a4b-it-q4

26B / 4B active / Q4

Min VRAM: 19 GiB

StandardBalanced quality

Gemma 4 26B A4B

google/gemma-4-26b-a4b-it

26B / 4B active

Min VRAM: 28 GiB

SmartestHighest quality

Gemma 4 31B

google/gemma-4-31b-it

31B params

Min VRAM: 36 GiB

Code & agents

Code generation and editing, and tool-calling autonomous agents

opencode, Kilo Code, OpenClaw

FastLowest latency

Qwen3.6 35B A3B Q4

qwen/qwen3.6-35b-a3b-q4

35B / 3B active / Q4

Min VRAM: 23 GiB

StandardBalanced quality

Qwen3.6 35B A3B

qwen/qwen3.6-35b-a3b

35B / 3B active

Min VRAM: 37 GiB

SmartestHighest quality

Qwen3.6 27B

qwen/qwen3.6-27b

27B params

Min VRAM: 31 GiB

Drop-in OpenAI compatible

No SDK to learn. Keep your existing OpenAI client — swap the base URL and key, and gigatoken handles routing to a provider device that serves your chosen model.

from openai import OpenAI

client = OpenAI(
    base_url="https://coordinator.gigatoken.baremetallabs.ai/v1",
    api_key="gt_...your inference key...",
)

response = client.chat.completions.create(
    model="qwen/qwen3.6-27b",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

How it works

Generate a key

Point your client

Set your OpenAI-compatible client base URL to the gigatoken endpoint.

Send requests

Call chat completions as usual — gigatoken routes to a device serving your model.

For providers

Have a GPU sitting idle? Put it to work.

Run the gigatoken daemon, serve catalog models, and join the provider pool.

Open the provider console