Send your training script and dataset. We provision the hardware, run the job to completion, and send back the weights, the curves and the logs. No credit card, no Colab timeouts, no 12-hour disconnects.
Bank transfer · FriMi · Koko | Starts at Rs 750 — you see real output before the next payment
| Colab free | Colab Pro | NevaraCloud | |
|---|---|---|---|
| Session limit | Disconnects, often under 4h | Up to 24h, not guaranteed | Runs to completion in tmux |
| GPU you get | T4 16GB, when available | A100 sometimes, luck-based | You choose: 4090 24GB or A100 80GB |
| Payment | Free | USD card required | LKR bank transfer or FriMi |
| If a dependency breaks | Your problem, at 3am | Your problem, at 3am | We debug it before the full run starts |
| Cost of an overnight idle session | Lost session | Burnt compute units | Nothing — we stop the box at completion |
Fill the form below with a link to your code and dataset — Google Drive, GitHub or Hugging Face. Tell us the framework, the base model and how many epochs you plan to run.
This proves your dependencies resolve, your dataset paths are correct and your code actually reaches the training loop. It also gives us a real per-epoch time, which is how we quote you accurately instead of guessing.
The sanity check costs Rs 750. It comes off your total if your code passes. See Terms §3.Measured epoch time × your epoch count, plus a 30% buffer. You approve it and pay before the full run begins. If the run later looks like it will exceed that buffer, we stop and ask you first — we never quietly bill you more.
Training runs inside a persistent terminal session on RunPod or Vast.ai, checkpointing as it goes. We watch for divergence, OOM kills and stalled epochs. The instance is terminated the moment the run finishes.
Weights, checkpoints, plots and the full training log, packaged and uploaded to your Drive. Downloadable for 14 days.
You send the code and data. We handle everything from CUDA versions to the final upload.
YOLOv8 / YOLOv11, image classification, CNNs. Typical run 2–4 hours on an RTX 4090.
Llama 3, Mistral, LoRA / QLoRA adapters on 7B–13B. Typical run 4–7 hours on an A100 80GB.
NevaraCloud.safetensors, merged model on requestDense segmentation, keypoint detection, Whisper and audio models. Typical run 3–6 hours.
Booking 14+ days before your deadline? 15% off. Referred someone who ran a paid job? Rs 1,500 off your next run.
Everything pre-installed — CUDA, PyTorch, the usual libraries. Open JupyterLab and start training. No setup, no version conflicts, no card required.
Hours count while the instance is running, not while you're typing. Shut it down when you step away.
.pt, .safetensors, .onnx on request.Most people are wary of transferring a large amount to someone they've never worked with. That's reasonable. So the first payment is small, and you get something real back before the next one.
We run one epoch of your code. If it fails, you get the full error output and our read on the cause — and that's where it ends, nothing further is charged. If it passes, you get a fixed quote and the Rs 750 comes off your total.
For jobs under Rs 12,000, the remaining balance is paid here and the job runs straight through to delivery.
Jobs of Rs 12,000 and above: half before the run starts, half once the training has finished and the artifacts are packaged and ready to hand over.
Koko lets you split any amount into three payments; the 4% covers their processing fee. Self-service passes are prepaid in full. We don't offer informal instalment arrangements — the staged payments above exist so you don't need one.
Clocks below start from your approval and payment, not from when you first message us.
| Self-service workspace provisioned | Within 4 hours, business hours |
| Sanity check and fixed quote | Within 24 hours of submission |
| Standard Vision delivered | 1–2 business days |
| Advanced Vision & Audio delivered | 2–3 business days |
| NLP & LLM fine-tune delivered | 2–3 business days |
| Research and multi-day runs | Quoted per job |
| Rush delivery — guaranteed 24 hours | +50% on the job price |
Capacity is finite and demand spikes around submission season. A date is only committed once we've confirmed it to you in writing — don't submit the night before and assume it lands.
Exact filenames depend on your framework. This is a representative YOLO run.
weights/best.ptBest checkpoint by val metricweights/last.ptFinal epoch, for resumingweights/best.onnxExported, on requestadapter_model.safetensorsLoRA adapters, fine-tune jobsresults.csvPer-epoch metrics, rawresults.pngLoss and metric curvesconfusion_matrix.pngNormalised and rawPR_curve.pngPrecision–recall by classtrain.logFull stdout, start to finishsamples/Inference outputs for eyeballingThese are the files your run produces. We package and hand them over — we don't write your analysis of them, and we don't decide what they mean for your project. That part is yours.
You get the full error output and our read on the likely cause, and nothing beyond the Rs 750 is charged. We don't debug your code as part of that — but you'll know exactly what broke and where. Fix it and resubmit, and the sanity check runs again.
No, not without you agreeing to it first. The quote is measured epoch time × your epoch count plus a 30% buffer. If a run looks like it will exceed that buffer, we stop and ask you before spending anything more.
Only us, and only for as long as the job needs. The training instance is terminated the moment the run finishes, and the Drive delivery link is retained for 14 days. For unpublished or confidential data, research and commercial jobs run on isolated secure-cloud hardware — say so when you submit.
That's a modelling outcome, not a failed run, so it's still billable. We deliver what your code produced — we don't tune your hyperparameters, rework your architecture or interpret your metrics. A run that completes with poor accuracy has done what you asked it to do.
Yes, within limits. Standard Vision and Advanced Vision & Audio include one free rerun after you fix your code; NLP & LLM fine-tunes include two. Beyond that, reruns are quoted as a fresh job.
Managed if you want the job handled end to end and priced per job, with idle time on us. Self-service if you already know your stack and would rather drive: you buy hours and get JupyterLab and SSH on a box we provision, everything pre-installed. If you're unsure, start managed — the Rs 750 sanity check is the cheapest way to find out where you stand.
Standard Vision lands in 1–2 business days, the other managed tiers in 2–3, timed from your approval and payment rather than your first message. Rush delivery guarantees 24 hours for +50%. Capacity is finite around submission season, so a date only counts once we've confirmed it to you in writing.
No. Bank transfer (Commercial, Sampath, BOC, HNB), FriMi, or Koko to split any amount into three payments for +4%. Bank transfer is free and the fastest to clear. Self-service passes are prepaid in full.
Two of them. Book 14 or more days before your deadline and 15% comes off the job price. Refer someone who goes on to run a paid job — they name you when they submit — and you get Rs 1,500 off your next run. There's no group rate: a project shared between four people costs the same as one person's, and you sort the split between yourselves.
We don't write your model code, clean or label your dataset, choose your architecture, or interpret your metrics and write up what they mean for your project. We run your training job on real hardware and hand back every artifact it produced. The research is yours.
Something not covered here? Message us on WhatsApp — reply within 24 hours. The full terms are in Terms of Service.
The more of this you fill in, the faster we can quote you. We reply within 24 hours, usually much sooner.