Models

There are no tiers. GenKaKu serves one text model to everyone, plus a browser-powered small model and a swarm model on the API, and a separate image model. All of them are uncensored.

Qwen3.8 27B Uncensored

• Cost: billed per token, about 1 credit for a typical message
• Model ids: qwen3.8-27b-uncensored, qwen3.8-27b-uncensored-think
• Runs on native workers via ollama — consumer GPUs, not a data center
• Uncensored — refusal behavior removed

This is what chat uses, and what c0mpute code runs on. There is nothing to select: every message goes to the same model, and every native worker on the network runs the same build.

What it can do:

• Web search and tools — the model decides when to search, calls the tool, and answers grounded in the results with citations. Through the API you can hand it your own functions the same way.
• Vision — send it an image and it reads it. (Image input. Making pictures is the image model below.)
• Thinking mode — extended chain-of-thought before it answers, for harder problems. There is no surcharge. Thinking tokens are output tokens, so a thinking answer costs more only because it is longer.

genkaku-pro

• Cost: billed per token, at the same rate as every other text model
• Model id: genkaku-pro
• Uncensored Qwen3.5, run by browser workers on WebGPU — 9B on a card that can hold it, 4B otherwise

The browser lane: a small, fast model on the widest supply in the network. Each worker serves the largest rung its GPU can hold, so a request is answered by the 9B or the 4B depending on who picks it up. It also serves free prompts. It can attempt tool calls but is less consistent at them than the 27B, so for agents use qwen3.8-27b-uncensored.

genkaku-swarm

• Cost: billed per token, at the same rate as every other text model
• Model id: genkaku-swarm
• MiniMax-M2.5 (229B), split across a swarm of contributor GPUs — no single machine holds the whole model

A 229B model running on hardware that could never hold it alone. Availability depends on a swarm ring being assembled and ready, so check the available flag from GET /v1/models before you depend on it.

Image generation

• Cost: 10 credits per image
• Chroma1-HD on dedicated image workers, uncensored

Available both as a tool the text model calls when you ask it for a picture, and as a direct endpoint. See Image generation.

What things cost

Text is billed per token. One rate card covers every text model, so the model you pick changes the answer, not the price.

	USD per 1M tokens
Input	$0.15
Output	$0.90

A credit is $0.001. Requests round up to whole credits with a floor of 1, so a typical message of about 1,200 tokens in and 600 out costs about 1 credit.

What	Credits
Typical message	about 1
Image	10

Length is the price. A short question costs less than a long one, and a long answer costs more than a short one. There is no separate charge for thinking.

Most people never pay per message, because a plan grants credits every day. Credits also come from topping up with USDC at 500 credits per dollar.

When you send a message we hold the most it could cost, then charge what it actually cost when the answer finishes and give the rest back. If a job fails or you disconnect, the whole hold is refunded.

What "uncensored" means

Corporate AI models (ChatGPT, Claude, Gemini) are trained to refuse certain topics. Ask about anything the company considers sensitive and you get a refusal. These aren't safety features — they're content policies imposed by corporations.

The uncensored builds GenKaKu runs have had this refusal training removed. They answer your questions directly without corporate-imposed restrictions.

Web search

Web search is model-driven. The model itself decides whether a question needs current information; when it does, the orchestrator runs the search (Brave Search API), extracts content from the top results, and hands it back to the model as a tool result. The model then answers grounded in real, up-to-date web content and cites its sources.

That means it can answer questions about recent events and look things up, instead of guessing from training data.