Training runs in dedicated clusters. Your data is never used to train other models.
Supervised fine-tuning, direct preference optimization, reinforcement learning from human feedback.
Fine-tune any model on Aigarth, including your own pre-trained checkpoints.
Multi-node, multi-GPU training. Horovod, DeepSpeed, FSDP supported.
Train, evaluate, and deploy in hours. Hot-reload checkpoints without service interruption.
The resulting model is published as an ANN you own, license, and earn from.
Token-efficient, with volume discounts and burn incentives.
Low-rank adaptation, fast
Full parameter fine-tuning
Preference optimization
Includes reward model training
Pricing is illustrative. Final rates are governed by on-chain parameters and may vary based on network state.
Tier-based access. Higher stakes unlock better economics and more capacity.
| Tier | Required stake | Access |
|---|---|---|
| Startup | 150M QUBIC | Fine-tuning enabled |
| Business | 500M QUBIC | Distributed, custom architectures |
| Enterprise | Custom | On-prem, private training |
Drop-in compatible with the OpenAI SDK.
from aigarth import Aigarth
client = Aigarth(api_key="sk-aigarth-...")
# Submit a fine-tuning job
job = client.fine_tuning.create(
model="aigarth-reason-1",
training_file="file-abc123",
method="lora",
hyperparameters={
"epochs": 3,
"lr": 2e-5,
"lora_r": 16,
},
)
# Wait for completion
model = job.wait()
print(f"Model published: {model.id}")Everything in the standard tier, plus the things enterprises need.