Pay for GPUs at cost. Pay us for the platform.
No per-token training markup, no surprise egress, no minimums. GPU time is passed through at the broker's price and billed per second. Bring your own compute and the GPU line goes to zero.
Free
For individuals learning post-training.
- 1 seat
- Models up to 8B
- SFT, DPO, KTO with LoRA
- Full eval suite
- 1 serverless endpoint
Build
For engineers shipping a first tuned model.
- Models up to 70B
- All methods incl. GRPO
- Multi-GPU, single node
- Data studio & synthesis
- 5 endpoints, multi-LoRA
Team
For ML teams running post-training as a practice.
- Any model size, multi-node
- Custom rewards & judges
- Judge calibration & release gates
- BYO cloud or Databricks
- SSO, audit logs
- Shared Slack channel
Enterprise
For regulated, large-scale and on-prem deployments.
- VPC or air-gapped on-prem
- Reserved GPU capacity
- SCIM, RBAC, data residency
- 99.9% uptime SLA
- Named post-training engineer
- Managed engagements
What will your run cost?
An honest estimate from sustained per-GPU training throughput, the compute multiplier of each method and current broker prices. Real runs vary with sequence length and packing; we'll give you an exact quote on your data.
- Estimated GPU cost
- —
- GPU-hours
- —
- Wall clock
- —
- Suggested cluster
- —
Platform fee is separate and flat. On your own GPUs, the cost above is what you'd pay your provider.
Assumes bf16, sequence packing, ~45% MFU. GRPO includes rollout generation. Read the full fine-tuning cost breakdown.
Serving fine-tuned models.
Fine-tuned adapters are served at base-model prices. No premium for being yours.
| Model class | Serverless input / 1M tok | Serverless output / 1M tok | Dedicated |
|---|---|---|---|
| Up to 8B | $0.10 | $0.20 | from $1.20 / GPU-h (L40S) |
| 9B – 35B | $0.30 | $0.60 | from $2.39 / GPU-h (H100) |
| 36B – 80B | $0.80 | $1.20 | from $2.39 / GPU-h (H100) |
| MoE up to 400B | $0.60 | $1.80 | from $3.19 / GPU-h (H200) |
No asterisks.
If something isn't covered here, ask us. We publish prices because we think you should be able to plan without a call.
How is GPU time billed?
Per second, at the rate quoted by the broker when your job is placed, with no minimums. You see the price cap you set, the price you actually paid and a per-run cost breakdown. On your own clusters or cloud accounts there is no GPU charge from POFE at all.
Is there a free tier?
Yes. The Explore plan is free forever for one seat, with $50 of GPU credit to start, small-model training up to 8B parameters, the full eval suite and one serverless endpoint. No credit card required.
What does a typical fine-tuning run cost?
A LoRA SFT run on an 8B model with 50 million training tokens costs roughly $10–$25 in GPU time. A full-weight DPO run on a 70B model is typically a few hundred dollars. GRPO is compute-heavier because of rollouts — budget 3–5× an SFT run on the same data. Use the calculator on this page for your own numbers.
Do you charge per token for inference?
Serverless endpoints are billed per million input and output tokens, with fine-tuned adapters at the same price as the base model. Dedicated endpoints are billed per GPU-hour with scale-to-zero, which is usually cheaper above about 20 requests per second.
Can we use our cloud commitments?
Yes. In VPC deployments, training and inference run in your own AWS, Google Cloud, Azure or Databricks account and draw down your existing commitments. POFE bills only the platform subscription.
What is included in a managed engagement?
A dedicated post-training team for a fixed scope: eval design, data strategy and synthesis, experiments across methods and model sizes, distillation, deployment, and a documented handover. Engagements are priced on scope and outcomes, not hours.
Need reserved capacity or on-prem?
Enterprise plans include committed GPU pricing, VPC or air-gapped deployment and a named post-training engineer.