What's on the platform
Six parts, one console and one bill. Most teams start with a dev node and grow into the rest.
Train and fine-tune
Fine-tune an open-weight model, re-train one you already have, or start from scratch. Training runs as a container job. Kick it off on demand or put it on a schedule, and it doesn't need a machine you sit and watch.
Training & fine-tuningServe on endpoints
Pre-built containers for vLLM, SGLang, Triton and Dynamo, or bring your own image. Endpoints are OpenAI-compatible and authenticated, they scale on request metrics, and they scale to zero when the traffic stops.
Inference & endpointsDev Nodes
Jupyter-style notebooks on NVIDIA GPUs, with pre-built images for PyTorch and Transformers, or your own image if you have one. Start a node in seconds, stop it when you're done, pay for the hours it ran.
AI dev nodesData and models
Object Storage for datasets, a Model Repository for weights and config files, and Container Registry for private images. When data loading is the bottleneck rather than the GPU, move the dataset onto the parallel file system.
AI object storageIntegrations
GitHub, Hugging Face and Weights & Biases connect directly, so code, models and run history stay where your team already keeps them.
GitHubHugging FaceWeights & BiasesTenancy and access
Dedicated clusters when you need the hardware to yourself, VPC attachment and private endpoints when traffic shouldn't touch the public internet, and identity and access management for teams who have to show an auditor who did what.
Certifications & compliance
Models you can deploy today
One click puts any of these behind an endpoint. Anything on Hugging Face or GitHub works too, and so does your own custom model once it's in the Model Repository.
Want tokens from an API and no GPU to manage at all? That’s Blaze. Same models, none of the deployment.
- DeepSeek R1
- DeepSeek V3
- Llama 4 Maverick
- Llama 4 Scout
- GPT-OSS 120B
- Hermes 3 405B
- Dolphin Mixtral 8x22B
- Sarvam-2B
- Any model from Hugging Face or GitHub, or your own custom model
From dataset to endpoint in four steps
None of them involve writing a Kubernetes manifest.
Explore
Start a dev node, mount your dataset from Object Storage, and try the idea against a catalog model before you spend anything on training.
Fine-tune
Point a job at your dataset and a base model. TIR provisions the GPUs, runs the container, and logs the run to Weights & Biases if that's where you track them.
Deploy
Promote the checkpoint to an endpoint. You get an authenticated HTTPS URL, an API token, and a playground to test in before you wire up the app.
Scale
Set the endpoint to scale on concurrent requests, requests per second, or a runtime metric like
vllm:num_requests_running. Change GPU or model later without rebuilding around it.
OpenAI-compatible, out of the box
Every endpoint gets a root URL; append the standard routes.
# illustrative — copy the exact URL from your endpoint's API Request tab
curl https://infer.e2enetworks.net/project/p-XXXX/endpoint/is-XXXX/v1/chat/completions \
-H "Authorization: Bearer $TIR_API_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-r1",
"messages": [{"role": "user", "content": "Summarise this contract clause:"}],
"max_tokens": 400
}'What teams build on it
A few of the workloads that run on TIR today, and why the shape of the work suits dedicated Indian capacity.
Payments
Transaction risk scoring inside the authorisation path, where the latency budget is measured in milliseconds. Routing and success-rate models. Dispute triage as an overnight batch on the same cluster.
Lending
Underwriting decisioning. Document intelligence over bank statements, ITR and GST filings. Vernacular collections agents with a hard escalation path to a human.
Insurance
Claims triage, document fraud detection and underwriting risk. The work is batch-shaped, which is exactly where reserved capacity costs less than paying by the hour.
Medical imaging
Radiology and pathology models that read scans and flag findings for a clinician to confirm. Patient data stays on Indian infrastructure, which is usually the first question a hospital asks.
Voice and support
Speech recognition and support agents that handle Indian languages and code-mixed speech. Endpoints absorb the evening call peak and scale back down overnight.
Analytics at speed
GPU-accelerated pipelines with Spark and RAPIDS, for teams whose nightly job stopped fitting in the night. Retail demand, pricing, and churn models sit here.
How it runs
The parts that decide whether a platform is pleasant to live with after month three.
Billing you can predict
Rates are published on the rate card, in rupees, with no quote to chase first. Dedicated capacity bills at a fixed rate, so a busy month and a quiet month cost the same.
Observability you already own
Request-level logs, latency percentiles and the distribution of model decisions, pulled into the monitoring stack you already run. No second dashboard nobody opens.
Support from engineers
The people who work on the platform are reachable directly, and they're in your timezone. The escalation path is published, so you know who to reach when something can't wait.
TIR, or your own stack
Rolling your own is the right answer for some teams. Here's the honest split.
TIR handles this for you
- GPU provisioning, drivers and images, so nodes and endpoints come up in seconds
- Autoscaling, health checks and rollouts for serving, including scale-to-zero
- Training orchestration, scheduled jobs, and weight downloads from Hugging Face or the Model Repository
- Compliance posture you inherit: SOC 2, ISO 27001, PCI DSS, and data that stays in India
Stay on raw GPUs if
- You already run a mature Kubernetes and serving stack, with a platform team behind it
- You need custom kernels, an unusual scheduler, or a network topology of your own design
- You want root on the node, which E2E's GPU cloud gives you on the same hardware
Either way the GPUs and the regions are the same. TIR is the layer above them.
Frequently asked questions
Everything teams ask before they start on TIR.
TIR is E2E Networks' AI platform. It gives you GPU dev nodes for experiments, container-based jobs for training and fine-tuning, and autoscaling endpoints for serving, plus the storage, model registry and access controls around them. Everything runs on NVIDIA GPUs in India.
Yes, and most teams do both in the same project. Train with notebooks, scripts or scheduled pipeline jobs, then deploy the result as an endpoint using a pre-built container or your own image. The checkpoint doesn't have to move anywhere in between.
The catalog covers DeepSeek R1 and V3, Llama 4 Maverick and Scout, GPT-OSS 120B, Hermes 3 405B, Dolphin Mixtral 8x22B and Sarvam-2B, each deployable in one click. Beyond the catalog you can point an endpoint at any Hugging Face or GitHub model, or serve your own custom model from the Model Repository.
Large enough for frontier open-weight models. A single B200 carries 192 GB of GPU memory and an H200 carries 141 GB, and multi-GPU configurations go well beyond a single node. E2E runs 1,024 B200s in one Indian cluster. Past a certain size it becomes a capacity conversation rather than a platform limit, so talk to us about the specific model.
You pay for the GPU capacity your nodes, jobs and endpoints consume, at the published INR rates, from ₹49/hr on an L4 up to ₹664.05/hr on a B200. Commit for a term and the rate drops substantially: a B200 goes to ₹290/hr, an H100 to ₹155.90/hr. Endpoints that scale to zero stop billing when idle. You can check the current pricing here.
In India, under Indian jurisdiction. E2E's GPUs run in two regions, Delhi NCR and Chennai, and your data stays on that infrastructure. Seven security and quality certifications cover the platform, including SOC 2, ISO 27001 and PCI DSS, and the reports are published rather than described. For workloads that shouldn't touch the public internet, endpoints can sit behind VPC attachment and private networking.
The API reference is at docs.e2enetworks.com/api/tir, and the platform guides and tutorials are at docs.e2enetworks.com/docs/tir. Each endpoint also has an API Request tab in the console with a ready-made cURL and Python sample for that exact deployment.
Ship your first endpoint this week
Start a dev node, fine-tune something small, and put it behind an endpoint before you decide on anything bigger. Our engineers will help you size the run.