Grid documentation
Guides for your first API call, model selection, training and enterprise evaluation. Grid access is by invitation.
Find the guide for your next step.
Use a model ID enabled for your account. Choose a task to find its guide.
Make your first API call
Get access, create a key and call an enabled model.
Choose a model and budget
Compare model access, candidates and the usage meter.
Adapt a model
Review training paths, datasets and the serving plan.
Evaluate an enterprise workflow
Define your baseline, integrations and acceptance criteria.
Prepare an enterprise purchase.
Confirm commercial terms, deployment scope and the evidence your reviewers need.
Serverless has no SLA. A cold lane can return 503 with Retry-After. Dedicated and Private terms are confirmed in writing.
Platform overview and operating limits
Grid is Forward Intelligence's platform for inference, routing and training. Start with an enabled model, then evaluate quality, response time and cost on your workload.
Inference uses an OpenAI-compatible API. Training has its own API and serving scope. Developer examples cover a first call with Python, Node.js or cURL.
Start here#
- Get started covers access, keys, enabled models and a first call.
- Concepts defines the words the rest of these pages use. Run mode, lane, pinning, cold start, and what the meter counts.
- OpenAI compatibility tells you what is supported today, so you can check a parameter before you depend on it.
The four run modes#
Tokens measure traffic across all four modes. Compare the workload's token usage, capacity commitments and included resources before moving it. The billing basis for Dedicated and Private is defined in the proposal.
| Mode | What it is for | Availability |
|---|---|---|
| Grid Serverless | Elastic traffic. Capacity follows demand and scales to zero when idle. | Available |
| Grid Batch | Asynchronous inference at a discount, in controlled windows. | Available |
| Grid Dedicated | Committed capacity for steady production traffic. | Talk to us |
| Grid Private | An isolated environment for regulated workloads. | Talk to us |
Run modes covers each one in full.
Train and route#
- Training: three paths. Custom training, Training API, managed jobs. Invitation.
- RL rollouts: sampling for teams that keep their own trainer.
- Connect: Grid in coding harnesses. Smart Router is the engine.
- Custom projects: a quoted workload we design and run with you on Grid. Not a training job.
What Grid does not do#
Being direct about this is faster for you than discovering it later.
- Pricing publishes rate card v1 references for eligible serving variants. Billing and metering explains tokens, cache, Batch and capacity commitments.
- There is no service level agreement on Serverless. Dedicated and Private are contractual, and those conversations are with us directly.
- Serverless scales to zero, so a cold lane can answer
503with aRetry-Afterheader. Keep your retry logic. Errors and cold start covers the behaviour. - There is no model marketplace. The catalog is curated, and that is the point.