Skip to content
Part of NVIDIA Inception

Train AI models on your data
in minutes, not weeks.

Fine-tune, evaluate, and serve open-weight models that beat general APIs at a fraction of the cost.

Fine-tune and serve
  • Qwen3
  • Qwen3.8
  • Llama 3.3
  • Gemma 4
  • DeepSeek
  • gpt-oss
  • Mistral
  • Qwen3-VL
  • Phi-3
  • ModernBERT

Why fine-tune

Why fine-tune instead of calling a bigger model?

Stop paying frontier token prices for repeatable work. Train a smaller model you own.

Before

General model API

Renting intelligence by the token

  • Long prompts billed again on every request
  • Silent vendor updates change answers overnight
  • Generic weights miss your schema and domain rules
  • Data leaves your cloud on vendor terms

After

Your fine-tuned model

Smaller, faster weights built for your task

  • Higher accuracy, proven on held-out test sets
  • Per-second GPUs that scale to $0 when idle
  • Runs on Tensorant compute or your own cloud
  • Standard LoRA weights you can export anytime

Accuracy

Verified before deploy

Compare tuned and baseline answers side by side on held-out splits.

Evaluations

Unit economics

Zero idle waste

Pay per second on active GPUs and $0 whenever traffic stops.

LLM inference

Throughput

Tuned for your GPU

Autoresearch benchmarks serving settings automatically to lower token cost.

Autoresearch

Ownership

No lock-in

Keep data private, run on six clouds, and download weights anytime.

Fine-tuning

Platform

Datasets, fine-tuning, evaluation and inference in one place

One shared project from raw documents to a live endpoint. Start with any stage.

Take the product tour

Use cases

Real production workloads that pay off fast

Focused, high-volume tasks where small fine-tuned models beat general APIs on quality and cost.

See every use case

Classify and route

Route tickets, triage requests, and flag escalations in milliseconds at low cost.

I was charged twice for March. Can you refund one of them?

  • billing
  • technical
  • account
  • shipping

Scored withaccuracy, per label

Classification and routing

Extract to JSON

Pull clean, typed JSON from invoices, forms, and emails on the first pass.

Invoice INV-2041

Acme Ltd

Due 14 November 2026

Total due $1,280.00

{
  "invoice": "INV-2041",  "vendor": "Acme Ltd",  "due": "2026-11-14",  "total": 1280.00}

Scored withJSON schema, field accuracy

JSON extraction

Answer from your documents

Turn your docs into training pairs and answer questions with cited sources.

How do I put the router back to factory settings?

Hold the reset button for ten seconds, until the light blinks amber.
router-guide.pdf, page 4

Scored withToken F1, or an LLM judge with your rubric

Fine-tuning vs RAG

Read images

Answer questions over charts, scans, and screenshots with a fine-tuned vision model.

Which quarter grew the most?

Q3, up about two thirds on Q2.

Scored withthe preset that fits the question

Vision models

GPUs

Use our GPUs, or connect your own cloud account

Add credits and pay per second on Tensorant compute, or connect your own cloud GPU account.

See GPU pricing
  • Zero surprise bills

    Every job previews its maximum GPU spend before you approve launch.

  • Hard hour and price caps

    Set runtime and hourly rate limits so jobs never overrun your budget.

  • Instant GPU release

    Machines shut down the second training or evaluation finishes.

  • $0 while traffic sleeps

    Wake-on-request endpoints scale to zero between requests and cost nothing overnight.

API

Swap three lines to ship faster, cheaper models

Published models speak the OpenAI chat completions API. Change three lines and ship.

API quickstart
OpenAI-compatible
import os
from openai import OpenAI
 
client = OpenAI(
    base_url="https://api.tensorant.io/v1",
    api_key=os.environ["TENSORANT_API_KEY"],
)
answer = client.chat.completions.create(
    model="support",
    messages=[{"role": "user", "content": "How do I reset my router?"}],
)
print(answer.choices[0].message.content)
# Hold the reset button for ten seconds, until the light blinks.

Let your coding agent run the workflow.

Connect Claude Code, Cursor, or Codex over MCP with scoped permissions and cost approval.

MCPhttps://mcp.tensorant.io/mcp

Claude Code, Cursor and Codex. On every organization.

Read
Looks at projects, data, runs and models. Changes nothing.
Build
Also adds and reviews data, freezes versions and runs evaluations.
Full
Also starts training and deployments, after showing the price.

Pricing

Pay as you go, with no subscription

Add credits and pay for GPU time, training tokens and storage as you use them. To train and serve on your own GPU accounts, subscribe to Own compute.

Pay as you go

$0/ month

No subscription. Pay from credits for what you use.

  • GPU time priced per hour, billed per second
  • Training-token fee from $0.10 per 1M tokens
  • 5 GB of storage included
  • Images, RL and decision models
  • Up to 10 members
  • Email support

Add credits from $10 when you're ready to train.

Own compute

$250/ month

$2,500/ year

Train and serve on your own GPU accounts.

  • Everything in Pay as you go
  • Your own GPU accounts on six supported clouds
  • Your provider bills GPU time; the training-token fee comes from credits
  • Priority email support and an onboarding call

Subscribe in the console. Yearly billing gives you two months free.

Enterprise

Custom

For larger teams, self-hosting or custom limits.

  • Self-hosted on your own infrastructure
  • Custom limits and members
  • Your own GPU accounts
  • Dedicated support

Tell us what your team needs.

FAQ

Frequently asked questions

Anything else? Write to us.

What is fine-tuning, and when is it worth it?

Fine-tuning trains an existing open model further on examples of your own task, so it answers the way your examples do. It is worth trying when a general model costs too much, or gets a narrow, repeated job wrong: classifying, extracting fields, answering from your documents. Tensorant scores the tuned model and the base model on the same test set, so you see whether it helped before you ship it. See fine-tuning vs RAG.

Do I need to write code?

No. Building the dataset, training, evaluating, running Inference Autoresearch and serving all happen in the console, or through your agent over MCP. You write code only to call your published models, with the OpenAI SDK or any HTTP client.

Do I need my own GPUs?

No. Tensorant compute gives you GPUs with nothing to set up: on Pay as you go you add credits in the console and pay per second for what runs. With the Own compute subscription, you can also use your own GPU accounts (RunPod, Verda, Lambda, AWS, Azure or Google Cloud) and the storage that comes with them, and your provider bills you for GPU time. Your own S3 storage is free on every organization. See GPU pricing.

What does it cost?

Pay as you go, with no subscription: you pay from credits for GPU time on Tensorant compute, a training-token fee from $0.10 per 1M tokens, and storage past the 5 GB included. With the Own compute subscription, $250 a month or $2,500 a year, you can also use your own GPU accounts and the storage that comes with them, and your provider bills GPU time. Your own S3 storage is free on every organization, and Enterprise is custom. Model endpoints you add to write or judge examples bill separately, and a sleeping model costs no GPU time. See pricing.

Which models can I fine-tune?

Public, ungated models on Hugging Face in the Qwen3, DeepSeek, Llama, Mistral, gpt-oss, Phi-3 and GPT-NeoX families, and vision models such as Gemma 4 and Qwen3.8. Every organization has images, RL and decision models. See supported models.

Can I train on preferences instead of answers?

Yes. DPO learns from pairs of a preferred and a less preferred answer to the same prompt, with text or image prompts and LoRA or QLoRA. See DPO preference training.

Where does my data live?

In Tensorant's storage, kept separately for your organization, unless you connect your own storage, and then it stays in your cloud. Requests to a published model pass through on their way to its GPU.

Can I take my model elsewhere?

Yes. Download the adapter as standard LoRA files that load with PEFT on the same base model, or publish it to your own Hugging Face account. See publishing to Hugging Face.

Start with one dataset.

Create an account and import your data, or book a demo and we will show you the product.