Skip to content
Stack

Bring your own stack.

Your models, your compute, your sandboxes, your trainer. Nenyax orchestrates them, and any piece swaps without touching the environment.

Connect your stack Read the docs
01Your environmentstays put02Modelany endpoint03Computeyour account04Sandboxany backend05Trainerthe one you use
The idea

Neutral by design.

You pay your providers

Model calls and GPUs are billed by the companies you already use, at their prices. Nenyax never resells either.

Your machines run the jobs

A runner on your laptop, GPU box or cloud does the work. The platform owns no compute.

Swap any piece

Every piece is chosen by name, so changing one never means rewriting the environment.

Compute

Your machines run the jobs.

A runner is a small process you start on any machine you control. It connects out to Nenyax, picks up your queued runs, and streams results back.

  • It dials outHeartbeat, claim, events, finish: outbound only. Nothing reaches in for your keys.
  • A laptop, a GPU box or your cloudStart it once and it picks up every run in your workspace.
  • Keys and data stay putModel keys and data never leave the machine the runner is on.
heartbeat01Your machinelaptop, GPU box or cloud02Keys stay heremodel keys and data03Dials outoutbound only04Nenyax owns no computeit only orchestrates
Every layer

One plug per layer.

A project's stack has a model, compute, a sandbox, a trainer and storage. Each is a choice, not a commitment.

Any OpenAI-compatible endpoint: hosted providers with your own key, or vLLM, SGLang, Ollama and mlx-lm on your machine. Every call is recorded.

Any OpenAI-compatible modelyour own key

Hosted

OpenAI Anthropic Gemini OpenRouter Together Fireworks

On your machine

vLLM SGLang Ollama mlx-lm
Integrations

81 integrations, honestly labelled.

Hubs, model providers, local models, compute, sandboxes, trainers, tracking and storage. Each one says whether it is ready or still in beta.

34 ready · 40 beta · 1 planned
Prime Intellect Hub
Hugging Face
OpenAI Evals
lm-eval-harness
reasoning-gym
Harbor
Gymnasium
GitHub
Gitee AI
Kaggle
Modelers
NVIDIA NGC
OpenXLab
Security

Built so your keys stay yours.

Your machine runner online

OPENAI_API_KEY HF_TOKEN WANDB_API_KEY
Outbound only:heartbeatclaimeventsfinish
Keys

Never stored by Nenyax

Hosted services use your own key, kept on the runner's machine. Only heartbeats, claims, events and results go back.

Isolation

As much as the task needs

From none to process, container or microVM, declared by the environment.

Plugins

Your own backends

A sandbox, provider or tracker you run can be added as a plugin with your key.

Langfuse LangSmith Phoenix Claude Code sessions
scrubbed→
Environment tasksSFT pairsDPO preferencesDistillation set
Traces

Scrubbed before upload

Coding-agent sessions are scrubbed of secrets and personal data by the CLI before anything leaves your machine.

Get connected

Three steps to a working stack.

1

Install

Install the Nenyax CLI from your workspace on the machine that will run jobs.

2

Start a runner

Create a token under Compute, then start the runner with it.

nenyax runner --token <token>
3

Connect a model

Add a provider key under Connections, or point at your own server.

Configure

Change a line, not the environment.

Every piece is chosen by name, in Python or in a config file. No user code imports a vendor class.

run.toml
[env]
uri = "reasoning_gym:basic_arithmetic?size=200"
[model] # any provider: vllm, ollama, openrouter, together, ...
provider = "vllm"
name = "Qwen/Qwen3-8B"
[learner] # incontext | grpo | tinker | my_pkg.learners:MyLearner
use = "incontext"
[train]
rounds = 10
group_size = 8

Swap the provider or the learner and run it again.

Questions

Good to know.

No. Nenyax owns no compute. Jobs run on machines you control, from a laptop to a cloud cluster.

Plug in what you have.

Connect a model and a machine, then evaluate or train on any environment in the hub.

Get started Read the docs