What is VoltageGPU API?

VoltageGPU provides two APIs: a GPU Pods API for deploying and managing GPU instances, and an AI Inference API for running 140+ AI models including DeepSeek-R1, Llama 3, and FLUX.

Is VoltageGPU API OpenAI compatible?

Yes, the VoltageGPU AI Inference API is fully OpenAI-compatible. You can use existing OpenAI SDKs by simply changing the base URL to https://api.voltagegpu.com/v1

What AI models are available?

VoltageGPU offers 140+ AI models including DeepSeek-R1, DeepSeek-V3, Llama 3.3 70B, Qwen 2.5, Mixtral, FLUX.1, Stable Diffusion, and many more for chat, embeddings, image and video generation.

✨Fine-Tuning API

Fine-Tuning API

Train custom AI models on decentralized GPUs. Powered by Bittensor Subnet 56 (Gradients).

Base URL: https://api.voltagegpu.com/api/volt/fine-tuning

Quick Start

Create your first fine-tuning job with a single API call:

curl -X POST "https://api.voltagegpu.com/api/volt/fine-tuning" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-llama-finetune",
    "taskType": "text",
    "model": "meta-llama/Llama-3.1-8B-Instruct",
    "dsRepo": "username/my-dataset",
    "hours": 2
  }'

Tip: Prepare your dataset on HuggingFace first. The API pulls directly from your HF repo.

Training Types

VoltageGPU supports four training methods, each suited to different use cases.

Text SFT (Supervised Fine-Tuning)

Train models on instruction/output pairs. Best for building custom instruction-following models.

Key Fields

`taskType`	required	Set to `"text"`
`fieldInstruction`	optional	Column name for instruction field (default: "instruction")
`fieldOutput`	optional	Column name for output field (default: "output")

// Dataset format (JSONL)
{"instruction": "Translate to French", "output": "Traduire en français"}
{"instruction": "Summarize this text", "output": "Here is a concise summary..."}

Chat SFT

Train on multi-turn conversation data. Ideal for building chatbots and conversational agents.

Key Fields

`taskType`	required	Set to `"chat"`
`chatTemplate`	optional	Template format: chatml, llama3, qwen3, mistral_v3_tekken, deepseek_v3, phi_4, gemma3
`fieldConversation`	optional	Column name for conversation field (default: "conversations")

// Dataset format (JSONL)
{"conversations": [{"role": "user", "content": "Hello"}, {"role": "assistant", "content": "Hi!"}]}
{"conversations": [{"role": "user", "content": "What is ML?"}, {"role": "assistant", "content": "Machine learning is..."}]}

DPO (Direct Preference Optimization)

Align models to human preferences using chosen/rejected response pairs. Best for RLHF-style alignment.

Key Fields

`taskType`	required	Set to `"dpo"`
`fieldPrompt`	optional	Column name for prompt field (default: "prompt")
`fieldChosen`	optional	Column name for preferred response (default: "chosen")
`fieldRejected`	optional	Column name for rejected response (default: "rejected")
`chatTemplate`	optional	Template format (same options as Chat SFT)

// Dataset format (JSONL)
{"prompt": "Write a poem", "chosen": "Roses are red, violets are blue...", "rejected": "I don't write poems"}
{"prompt": "Explain gravity", "chosen": "Gravity is a fundamental force...", "rejected": "Things fall down"}

GRPO (Group Relative Policy Optimization)

Train with reward functions instead of human-labeled preference data. The model learns by comparing its own outputs within a group, guided by custom reward signals.

Key Fields

`taskType`	required	Set to `"grpo"`
`fieldPrompt`	optional	Column name for prompt field (default: "prompt")
`rewardFuncs`	optional	Array of reward function names to use during training
`chatTemplate`	optional	Template format (same options as Chat SFT)

curl -X POST "https://api.voltagegpu.com/api/volt/fine-tuning" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-grpo-alignment",
    "taskType": "grpo",
    "model": "meta-llama/Llama-3.1-8B-Instruct",
    "dsRepo": "username/my-prompts",
    "hours": 3,
    "chatTemplate": "llama3"
  }'

Custom Dataset (Chat & Text)

Upload your dataset as a direct URL instead of a HuggingFace repo. Supports both chat and text SFT formats.

Key Fields

`taskType`	required	Set to `"custom_chat"` or `"custom_text"`
`datasetUrl`	required	Direct URL to your dataset file (JSONL, CSV, or JSON)

curl -X POST "https://api.voltagegpu.com/api/volt/fine-tuning" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "custom-dataset-train",
    "taskType": "custom_chat",
    "model": "Qwen/Qwen2.5-7B-Instruct",
    "datasetUrl": "https://my-server.com/data/conversations.jsonl",
    "hours": 2,
    "chatTemplate": "qwen3"
  }'

Image LoRA

Fine-tune FLUX and SDXL diffusion models with LoRA adapters. Best for custom image styles, brand assets, and characters.

Key Fields

`taskType`	required	Set to `"image"`
`imageModelType`	optional	Diffusion model type: `"flux"` or `"sdxl"` (default: "flux")

Dataset: Upload your images to a HuggingFace dataset repo. Include a metadata file mapping image filenames to captions.

Pricing

85% markup on Gradients (SN56) base rates. Charged upfront for the full job duration.

Model Size	Price / Hour	Example Models
Small (1-3B)	$18.50	Phi-4-mini, SmolLM
Medium (7-13B)	$27.75	Llama-3.1-8B, Qwen2.5-7B, Mistral-7B
Large (30-70B)	$46.25	Llama-3.1-70B
Image LoRA	$18.50	FLUX, SDXL

API Endpoints

List Jobs

Retrieve all your fine-tuning jobs, optionally filtered by status.

Method	Endpoint	Description	Auth
GET	`/api/volt/fine-tuning`	List all fine-tuning jobs	X-API-Key

Query Parameters

status optional Filter by status: pending, training, evaluating, completed, failed

Create Job

Start a new fine-tuning job.

Method	Endpoint	Description	Auth
POST	`/api/volt/fine-tuning`	Create a fine-tuning job	X-API-Key

Request Body Parameters

`name`	required	Job name (alphanumeric, hyphens allowed)
`taskType`	required	Training type: text, chat, dpo, grpo, image, custom_chat, or custom_text
`model`	required	Base model ID (e.g., "meta-llama/Llama-3.1-8B-Instruct")
`dsRepo`	required	HuggingFace dataset repo (e.g., "username/my-dataset")
`hours`	required	Training duration in hours (1-24)
`fieldInstruction`	optional	Instruction column name (text SFT)
`fieldOutput`	optional	Output column name (text SFT)
`chatTemplate`	optional	Chat template format (chat SFT / DPO)
`fieldConversation`	optional	Conversation column name (chat SFT)
`fieldPrompt`	optional	Prompt column name (DPO)
`fieldChosen`	optional	Chosen response column name (DPO)
`fieldRejected`	optional	Rejected response column name (DPO)
`imageModelType`	optional	Image model type: flux or sdxl (image LoRA)

Get Job Status

Retrieve the current status and details of a specific job.

Method	Endpoint	Description	Auth
GET	`/api/volt/fine-tuning/:id`	Get job status and details	X-API-Key

Rename Result Model

Rename the output model after a successful training run.

Method	Endpoint	Description	Auth
PATCH	`/api/volt/fine-tuning/:id`	Rename model or backup repo	X-API-Key

# Rename the result model
curl -X PATCH "https://api.voltagegpu.com/api/volt/fine-tuning/JOB_ID" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"action": "rename", "modelName": "my-custom-llama-v2"}'

# Backup training repo
curl -X PATCH "https://api.voltagegpu.com/api/volt/fine-tuning/JOB_ID" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"action": "backup"}'

Cancel Job

Cancel a running or pending fine-tuning job. Jobs cancelled before training starts are automatically refunded.

Method	Endpoint	Description	Auth
DELETE	`/api/volt/fine-tuning/:id`	Cancel a fine-tuning job	X-API-Key

Supported Base Models

Choose from popular open-source models as your fine-tuning base:

Language Models

meta-llama/Llama-3.1-8B-Instruct
meta-llama/Llama-3.1-70B-Instruct
Qwen/Qwen2.5-7B-Instruct
mistralai/Mistral-7B-Instruct-v0.3

Image Models

FLUX — Black Forest Labs diffusion model
SDXL — Stable Diffusion XL

Workflow

From dataset to deployed model in four steps:

Prepare Dataset

Upload your training data to a HuggingFace dataset repository in the correct format for your chosen task type.

Create Job

Submit a fine-tuning job via the API or dashboard. Choose your base model, task type, and training duration.

Monitor Progress

Track your job through its lifecycle: pending → training → evaluating → completed

Deploy or Download

Once complete, download the trained weights from HuggingFace or deploy directly via the VoltageGPU API.

Resources

✨

Fine-Tuning API

Quick Start

Training Types

Text SFT (Supervised Fine-Tuning)

Key Fields

Chat SFT

Key Fields

DPO (Direct Preference Optimization)

Key Fields

GRPO (Group Relative Policy Optimization)

Key Fields

Custom Dataset (Chat & Text)

Key Fields

Image LoRA

Key Fields

Pricing

API Endpoints

List Jobs

Query Parameters

Create Job

Request Body Parameters

Get Job Status

Rename Result Model

Cancel Job

Supported Base Models

Language Models

Image Models

Workflow

Prepare Dataset

Create Job

Monitor Progress

Deploy or Download

Resources

Dashboard

Showcase

Discord

Support