Skip to content
Tensovy
How it works

One sentence in. A trained model, a registered run and your own weights out.

Below is the complete Tensovy pipeline. Every stage runs inside one agent loop, on Tensovy managed infrastructure or your own cloud account, with no manual file shuffling between steps.

1MODEL DISCOVERY AND CONSULTATION

Tell the agent the outcome, not the hyperparameters

You open with a goal in plain language. The agent turns that into an architecture decision, a memory budget and a compute plan before a single dollar of GPU time is spent.

  • Base model recommendation matched to task type and output schema
  • VRAM estimate calculated from sequence length, batch size and adapter rank
  • GPU tier suggestion with the live hourly rate for the route you pick
  • A cost and duration estimate you approve before anything launches
agent consultation
> extract structured line items from messy invoices
task vision language, structured output
base Qwen2.5-VL-7B-Instruct
method LoRA r=16, alpha=32, bf16
vram ~21 GB estimated
gpu 1x RTX 4090 (24 GB)
est. ~48 min, about $0.35 total
2DATA PIPELINE AND DATASET STUDIO

Bring a labelled dataset, or find one

The dataset stage is where most fine tuning projects die. Tensovy takes the labelled data you already have, or helps you find a public dataset when you have almost none, and previews and profiles every row so you see exactly what the trainer will see before a GPU starts.

  • Upload of labelled JSONL, CSV, an image folder or a zip of one
  • Public connectors that search and import Hugging Face and Kaggle datasets
  • Row preview and a dataset profile so you see what the trainer will see before it runs
  • A data loader written for your recipe inside the generated training package, so the trainer reads your dataset as it is
dataset studio
scanning uploads/ 1,000 files
detected invoice scans, line item labels
preview 20 rows rendered
profile 1,000 rows, labels present
recipe Qwen2.5-VL, chat template
staged object storage, your account
artifact dataset-8fk2
3INFRASTRUCTURE PROVISIONING

Managed, or your own cloud

Run training on Tensovy managed infrastructure and get one bill, or connect your own cloud account and let your provider bill you directly with nothing added by us. Same agent, same workflow, same weights either way. Both routes are in early access.

  • Managed by Tensovy, with no cloud account to set up and one bill
  • Or your own cloud account, with the key held server side and used only for infrastructure operations
  • Live inventory, hardware plan and hourly price cap shown for approval, and the instance starts only when you say yes
  • Training package and data pushed over SSH, no manual scp or volume wiring
provisioning
querying gpu inventory ...
region EU-RO-1 available
gpu RTX 4090 $0.44/hr
plan 1x RTX 4090, approved by you
push package + data over ssh
instance running in 51s
billing managed by tensovy, or your own cloud account
4SELF HEALING TRAINING LOOP

Failures get fixed, not forwarded to you

Before the long run, the agent smoke tests the package on the GPU. When it breaks, the agent reads the traceback, classifies the failure and applies the correction that an experienced engineer would apply, then reruns until it passes. Only then does training launch and stream live to your dashboard.

  • Live loss curves, log tail, VRAM headroom and running spend
  • Dependency pins, API changes and tokenizer errors fixed by the agent, not by you
  • Dry run passes before the long run, so GPU hours go to training instead of debugging
  • Hourly price cap guard and a cost meter in view, so a run cannot quietly burn budget
training logs
[04:12:07] smoke test: python train.py --dry-run
[04:12:19] ImportError: cannot import name 'SwizzleType' from 'torchao'
[04:12:19] agent: traceback classified. pinning torchao<0.17
[04:13:41] agent: rerun passed. loss 0.213 vram 7.9/24.0
[04:13:44] launch approved. step 1 loss 1.842
[04:41:33] complete. best val loss 0.211
[04:41:34] weights pulled to storage. shutdown is one click
5EXPORT AND WEIGHT DELIVERY

A registry entry you can trace and a file you can keep

When the run finishes, the weights are pulled off the instance into Tensovy object storage under your account and recorded in your model registry. Shutting the instance down is one click, and the raw artifacts stay yours.

  • One click shutdown of the instance with the meter in view, so you decide when billing stops
  • Adapter or merged model recorded in the registry with its config, dataset and metrics
  • Roadmap: an OpenAI compatible endpoint served from your registered weights
  • Direct download of safetensors weights, plus PyTorch weights for detection models
export
merging adapter ... done
pulling weights from instance ... done
recording run in registry ... done
registry
runs / 8fk2 val loss 0.211 Qwen2.5-VL-7B
download adapter.safetensors 268 MB
Under the hood

Standard tooling, generated for you instead of by you.

Tensovy does not invent a proprietary trainer. It writes configs for the libraries the ecosystem already trusts, which means every run is reproducible outside the product.

  • Unsloth

    Memory efficient LoRA and QLoRA fine tuning for language models

  • LibreYOLO

    YOLO family detection training on your own labelled images

  • Hugging Face

    Model hub and dataset search, imported from inside the workspace

  • Kaggle

    Public dataset search and import when you have almost no data

  • PyTorch

    The runtime underneath every generated training package

  • RunPod

    The GPU cloud behind bring your own key today, alongside Tensovy managed infrastructure

Run the whole pipeline yourself.

Early access is open now. A task is all you need to start, and bringing your own cloud account is optional.