What is Zils?
Zils helps applications choose between defined outcomes using text or images. You supply the context and possible answers. A decision model returns probabilities, and your application decides what to do with them.
Use Zils at a decision point in an existing product: route a support request, select an agent's next tool, or classify a product photo. These are tasks to evaluate on your own data, not promises of accuracy. Your application keeps control of rules, permissions, actions and human review.
Example: route a support request
Input: “Please send the invoice for my last order.”
Question: Which team should handle this request?
Possible result: billing 90% · technical 6% · account 4%
Application action: route to billing if your tested review policy allows it; otherwise ask a person to review.
These probabilities are illustrative. They are not a live model result or a correctness guarantee.
Read how Zils fits into your application for the complete workflow, or how decision models work for the concepts.
Choose your starting point
How the pieces fit together
Developers call decision models and supply examples when they want a model tuned to their task. On the Bittensor network, miners build better base models: each size of the z3 line is a soup pool, and miners commit model bodies that validators average into the pool's next release. Miners also train customers' open-tier models. The platform runs accounts, billing, training jobs and hosted decisions.
You can start with an available model without training anything. An adapter is a set of trained changes that specializes a base model for your task. New text and image adapters use z3-mini, whose founding baseline is the unchanged d3-mini checkpoint adopted on 17 October 2026. Training produces a candidate to evaluate, and may end with no qualifying model. The model guide explains selection and compatibility.
What works today
| Start with | How |
|---|---|
| An account | Get started with Google or an email link; anyone can sign up |
| Hosted decisions | An API key; see the quickstart |
| Training | On the website, the CLI, a coding agent over MCP, or zils fit train; open tier free, private $0.20 per million tokens. See the training workflow |
| Price | $0.06 per million input tokens, output free. Prepaid top-ups start at $5. See pricing and limits |
| SDKs and CLI | npm install -g @zils/cli and npm install @zils/sdk (TypeScript, 0.4.18); the Python SDK is available from source. See SDKs and CLI |
| Find where Zils fits | npx -y @zils/cli fit agent in your project, or zils fit |
| Mining | Commit a model body to a soup pool with zils-miner, installed by curl -fsSL https://zils.ai/install.sh | sh, or train customer jobs |
| Validating | See Run a validator |
Hosted decisions run on platform GPUs today; serving from miners is not live yet. Read what's available for the boundary between current features and planned work.
For questions or a pilot, contact Zils. To contribute code or documentation, use the repository map.