Per-token credits. Public rates. No seats, no minimums.
One credit equals one US dollar of billed compute. Buy credit packs outright or subscribe monthly — the per-token rate is the same. Adapters are always yours to download.
Buy once. Spend whenever.
Larger packs include a volume bonus. Packs never expire.
Predictable monthly allowance.
Around 20% more credits per dollar than pay-as-you-go packs, in exchange for a recurring commit. Top up with packs any time.
Small, steady usage
- 1 concurrent training job
- 7-day job history retention
- Adapter download on every run
- Top up any time with credit packs
Regular fine-tuning cadence
- 3 concurrent training jobs
- 30-day job history retention
- Priority queue placement
- Top up any time with credit packs
Teams running many jobs
- 10 concurrent training jobs
- 90-day job history retention
- Dedicated support
- Top up any time with credit packs
Per million training tokens.
You’re billed on every token the trainer sees — rows × context × epochs. Failed runs (pre-training) are fully refunded.
| Model size | Price / 1M tokens | Credits / 1M tokens |
|---|---|---|
| ≤ 3B | $0.11 | 0.11 cr |
| 7 – 8B | $0.59 | 0.59 cr |
| 13 – 14B | $0.99 | 0.99 cr |
| 32B | $3.32 | 3.32 cr |
| 70B | $8.30 | 8.30 cr |
Per million inference tokens.
Billed on prompt_tokens + completion_tokens from each response. Served via vLLM on shared serverless capacity.
| Model size | Price / 1M tokens | Credits / 1M tokens |
|---|---|---|
| ≤ 3B | $0.012 | 0.012 cr |
| 7 – 8B | $0.060 | 0.060 cr |
| 13 – 14B | $0.099 | 0.099 cr |
| 32B | $0.332 | 0.332 cr |
| 70B | $0.830 | 0.830 cr |
Hosted inference launches after the current training-focused alpha. During alpha you can download any adapter and serve it yourself from the first completed run.
Questions we hear.
One credit equals one US dollar of billed compute. Credit packs and subscription allowances are denominated in credits; the billing ledger shows every deduction at per-token granularity.
Nothing, if it fails before the first gradient step (infra error, bad dataset). Reserved credits are refunded in full. If training has already started, GPU time was consumed and credits are not refunded — partial results may still be downloadable.
Credit packs do not expire. Monthly subscription allowances do not roll over — unused credits reset on each renewal.
Yes. Subscription credits are consumed first each month; pack credits are used after the monthly allowance runs out.
No. Never. Your corpus and inference traffic are never used to train our base models. That's contracted in the DPA, not just promised.
Yes. Every completed job produces a downloadable adapter you can serve yourself. No lock-in.
Job submission is rejected with a 402 response. Top up a credit pack or upgrade your subscription and resubmit.