OpenAI /v1
compatible API. Your existing SDK works with a URL change
LoRA
adapters by default. ~10 MB artifacts, downloadable anytime
Any HF model
open causal LMs up to ~32B, priced by size tier
$0 idle
serving scales from zero and closes after 30 idle minutes
Everything between a dataset and an endpoint
The UI is a visual builder for the API. Anything you click, you can script.
On-demand GPUs
Each job runs on a dedicated GPU provisioned just for it. RTX 3090/4090/5090-class for smaller models, 80 GB datacenter cards above 16B. Spun up on demand, closed when the work ends.
OpenAI-compatible API
Files, fine-tuning jobs, chat completions. The same request and response shapes as OpenAI. Point your existing SDK at api.volition.network and it works.
LoRA, SFT and DPO
Supervised fine-tuning and preference tuning on curated Qwen3 bases or any open Hugging Face causal LM up to ~32B. Adapters are ~10 MB and download as tarballs.
Live training events
Per-step loss and DPO reward margins stream to the dashboard and the events API. Checkpoints save as you go; interrupted jobs resume from the last one.
Scale-from-zero serving
Finished models are servable immediately: vLLM nodes spin up on first request, share a GPU across adapters, and close after 30 idle minutes.
Metered usage & spend limits
Token-level metering on training and inference, daily rollups, and per-org spend limits that stop runaway costs before they happen.
Models & pricing
Prices are USD per million tokens. Fine-tuned adapters serve at the same price as their base model, and there are no charges while nothing is running. Fine-tuning jobs have a $2 minimum fee.
| Model | Context | Training | Input | Output |
|---|---|---|---|---|
| up to 5B | model's own | $0.25 | $0.06 | $0.10 |
| 5–9B | model's own | $0.35 | $0.12 | $0.20 |
| 9–16B | model's own | $0.45 | $0.18 | $0.35 |
| 16–35B | model's own | $0.90 | $0.40 | $0.70 |
Watch every step, terminate anything
Per-step loss and training events stream to the dashboard and the API in real time. Cancel any job with one call; failed runs keep their full event history so you can see exactly why.
trainingprovisioningfailed
ftjob-8yQmVt2pTRAININGloss 1.204
train lossstep 1,240
Contact us
Questions about pricing, larger models, or running Volition for your team? Send us a message and we'll get back to you.
