# Prime Intellect Adapter

**Interface:** `prime` CLI (Python). **Kind:** decentralized marketplace. **Dynamic per-cluster pricing, spot available.**

## Install & auth

```bash
pip install prime          # verify package name at docs.primeintellect.ai — CLI surface changes
prime --help               # auth check + discover current command surface
```

## Provision

The marketplace model: search/bid on clusters, take an ask, or submit parallel
bids for options. Exact command surface changes — always discover first:

```bash
prime --help
prime <search|rent|cluster> --help    # whatever the current surface exposes
```

Pattern (verify against current CLI):
1. List available clusters with pricing (on-demand and spot).
2. Either take the ask on a fitting cluster (GPU class per model-deployment
   skill) or bid below ask for spot.
3. On fill, you get SSH access to the node(s).
4. Install runtime + weights, health-check, `gpu_register` as usual.

## Destroy

Release/stop the rental via the CLI or console. Spot can be reclaimed by the
host at any time — treat spot jobs as retryable.

## Gotchas

- **Same GPU, 6x price spread across clusters** (H200 from $0.47 to $3.14/hr in
  the snapshot). Always shop per batch, never assume a provider-level rate.
- Spot H100 was ~60% under on-demand in the snapshot — the right lane for
  retryable video batches.
- CLI changes fast: if a command fails, check `docs.primeintellect.ai` before
  improvising.
- Parallel bids (H100/H200/B200/B300 options) are the intended way to run a
  batch — take the first fill.

## Reference

- Docs: https://docs.primeintellect.ai (CLI, marketplace, spot)
- Pricing examples: docs/pricing/prime.md