Fine-tune Nemotron with LoRA on your own data and publish it to the store as an OpenAI-compatible API. No GPUs to set up. Each time someone calls your model, the tokens it serves become your earnings.
Base model
Nemotron 3 Nano 4B
+ your data
LoRA training
Serverless AI Job
Endpoint
Serverless AI Endpoint
API calls
Earnings per token
Wednesday open?
Base Nemotron · 1 / 5
We are open Wednesday, from 11 AM to 3 PM!
Trained demo model · 5 / 5
No, the bakery is closed every Wednesday, including public holidays.
Upload a chat-format JSONL, pick a Nemotron base model and start training. You pay for the GPU time the run actually used.
Only have documents? Nemotron drafts training examples from them.
Your LoRA is loaded onto the base model in a vLLM endpoint. It stops on its own after 15 minutes without calls.
Check real responses in the in-browser playground.
Nemotron sends the same prompts to the base model and to your fine-tuned model, and compares the answers.
Set prices per 1M input and output tokens and per request, then publish to the store.
Users call your model at /v1 with their own API key. Each call credits 80% of your price to your earnings.
| Role | Nebius service | What it does |
|---|---|---|
| LoRA training | Serverless AI Jobs | A PEFT + TRL container trains the LoRA on one GPU. |
| Inference | Serverless AI Endpoints | vLLM loads the base model with your LoRA and answers calls. Stops when idle. |
| Storage | Object Storage | Holds datasets and trained LoRA adapters (S3-compatible). |
| Data and evaluation | Token Factory | Calls Nemotron as-is to draft training data from your documents and to evaluate the fine-tuned model. |
Nemotron 3 Nano 4BDefault
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
Nemotron 3.5 Lightning 30B-A3B
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
Point base_url at Forkline and set model to the model ID shown in the store.
from openai import OpenAI
client = OpenAI(base_url="https://forkline.ai/v1", api_key="fk_...")
res = client.chat.completions.create(
model="your-model",
messages=[{"role": "user", "content": "How do I return an item?"}],
)
print(res.choices[0].message.content)HTTP/1.1 200 OK
X-Forkline-Cost-JPY: …
{"model": "your-model", "choices": [ … ], "usage": { … }}Nemotron 3 Nano 4B / Nemotron 3.5 Lightning 30B-A3B. The default is Nemotron 3 Nano 4B.
The person who builds the model. For training, the estimate is held up front and settled on the actual time. Endpoints are billed per minute while they run.
They stop after 15 minutes without calls, and no GPU time is billed. A call to a stopped endpoint starts it and returns 503 with Retry-After.
Users buy prepaid credits, and each call deducts the model's price. Each user has their own API keys and balance.
Yes. Set base_url to https://forkline.ai/v1 and model to the model ID. Streaming is supported.
No. Datasets and trained LoRA adapters are kept in Nebius Object Storage until you delete the model or your account.
Publish a Nemotron trained on your data
https://forkline.ai/v1