Bases that are actually yours to keep.
Every base is open-weight. Train it, export it, run it on your own hardware. No licence that expires when the invoice does, and no clause that stops you serving the result.
| Base | Context | Tune time (24 GPUs) | From | Export weights | Best at |
|---|---|---|---|---|---|
| Penstock 8B | 128K | ~4 h | $180 | Yes | Retrieval, classification, structured extraction |
| Penstock 34B | 256K | ~14 h | $740 | Yes | Reasoning-heavy work and long documents |
| Diffusion XL | — | ~6 h | $310 | Yes | Product and brand imagery from a house style |
| Penstock 1.5B | 32K | ~40 min | $38 | Yes | Edge and on-device, where latency beats depth |
| Bring your own | varies | varies | metered | Yes | Any open-weight checkpoint that loads in transformers |
Choosing between them
Start at 8B
Most tasks that read like "understand our documents and answer questions about them" are solved by the 8B. It is cheap enough to run twice and compare, which is worth more than getting the size right first time.
Move up when the eval says so
Go to 34B when the held-out evaluation shows the 8B failing on reasoning rather than on recall. Longer documents alone are not a reason — the 8B already carries 128K.
Bring your own if you have one
If it loads in transformers, the scheduler will pick it up. You keep whatever
licence came with it; we do not add one.