Autonomous AI infrastructure.
Get the most out of your compute.
Tensile is the platform for running AI on the clusters you already have: training, inference, agents, RL and evals in one place. Niko, our infra agents, operate it with your team: they debug failed runs and keep your GPUs busy.
Everything your AI infra team runs, in one place.
One sign-in for your organisation. Researchers, AI engineers and DevOps share the same runs, findings and policies.
Connect
Point Tensile at your clusters and cloud accounts. It maps your workloads, GPUs, spend and SLOs.
Niko
Infra agents that debug failed runs, raise GPU utilisation and open tested fixes as PRs.
Research
Auto Research, training runs, evals and benchmarks for your researchers.
Deploy
Inference endpoints, canary rollouts and feature flags.
Observe
Live events and metrics, agent traces and service health.
Sandbox
An isolated VM per run for agents, evals and RL rollouts.
Registries
Versioned models, datasets and evals, so every result traces back to its inputs.
Cost and audit
GPU, CPU and token spend per workload, and a log of every action taken.
Infra agents that work alongside your team.
Niko connects to your organisation and learns how you run: models, traffic, SLOs and constraints. Then it works around the clock.
Debug failed runs
Reads logs, traces, scheduler state, GPU health and recent changes, and names the root cause with the evidence.
Raise GPU utilisation
Finds idle allocated GPUs, jobs queued while capacity sits free, and steps waiting on data or NCCL.
Test before it changes anything
Tries each fix as a controlled experiment and checks correctness, speed and cost against your bar.
Ship fixes as pull requests
Changes land as PRs to your Terraform, Helm or Slurm config. Your team reviews and merges.
Average GPU utilisation goes up.
Time to debug goes down.
Training, inference, agents, RL and evals in one API and one UI.
Post-train on your own cluster
SFT, DPO, PPO and GRPO++: any framework, any model, any training backend. The Tensile operator gang-schedules the job onto your Kubernetes/Volcano or Slurm cluster, streams live metrics and returns the trained adapter.
- SFT · full fine-tune · LoRA
- RL & DPO pipelines
- Multi-node gang-scheduling
- Parallel hyperparameter sweeps
from tensile import Workload # Gang-scheduled onto your K8s/Volcano or Slurm clusterjob = Workload.training( base="meta-llama/Llama-3.1-8B", method="lora", # sft · rl · dpo · lora dataset="s3://org/gold/v1", backend="volcano", # volcano · slurm gpus=8,).submit() job.stream() # live loss + metrics → /observabilityOne isolated VM per run, resumed in about 16 ms.
Each run gets its own hardware-isolated VM, on CPU or GPU. Network egress is limited to an allowlist set before the VM starts. A sandbox can be paused, checkpointed, resumed in about 16 ms or forked into many copies, so an agent keeps its memory and files. Use it for agents, evals and RL rollouts in place of a hosted sandbox such as Daytona, Modal or E2B. Training and large inference jobs run directly on your GPU clusters instead.
Tensile runs these workloads on Sandbox for you. Tensile Sandbox is also its own product, with its own console and SDK, for teams that only need sandboxes.
Engine numbers measured on bare-metal Intel; cost vs Daytona at 10,000 average concurrent sandboxes on on-demand GCP. See the full comparison →
Use the GPUs you already have.
Tensile runs on Kubernetes, Slurm and bare metal, across clouds, neoclouds and on-prem. You keep your accounts, your data and your frameworks. We don't sell compute; we work with teams that already have it.
From first connection to first fix.
Connect your infra
Point Tensile at your clusters and cloud accounts. Niko maps your workloads, GPUs, spend and SLOs. Credentials stay on your side.
Get a baseline
Your GPU utilisation, idle GPU-hours and the failures that cost you most, ranked.
Merge the first fix
Niko opens a tested pull request. You review it and decide.
curl -fsSL https://tensile-sandbox-api.researchcommons.ai/install.sh | sh