Skip to content
NewFree to start, no card needed

One loop for every RL environment.

Bring your own environments

Turn your traces and codebases into benchmarks and RL environments, evaluate and train any model on them with the stack you already pay for, and share what you make. Nenyax orchestrates the loop; every piece stays yours.

Start a project Explore the hub
01Any environmentany SDK, hub, trace or codebase02One contractNenyax runs it on your stack03Evaluateany model, any benchmark04Trainkept only if it improves05Share & earnin the hub, under your name

Loads every format, imports your traces, runs on the stack you already have

VerifiersOpenEnvNeMo GymHarborGymnasiumreasoning-gymHugging FacePrime Intellect HubLangfuseLangSmithClaude CodeVerifiersOpenEnvNeMo GymHarborGymnasiumreasoning-gymHugging FacePrime Intellect HubLangfuseLangSmithClaude Code
vLLMSGLangOllamaOpenRouterOpenAIAnthropicModalE2BRunPodKubernetesverlWeights & BiasesvLLMSGLangOllamaOpenRouterOpenAIAnthropicModalE2BRunPodKubernetesverlWeights & Biases
New to RL?

It's one loop.

A task, your model's try, a checker's score, then compare or train. Everything on Nenyax is this loop, repeated.

01A taskfrom an environment02Your modeltries it03A checkerscores the try04You compareor train on it, and repeat
How it works

Make it, measure it, improve it, share it.

Four steps, each with its own page. Start anywhere.

Langfuse LangSmith Phoenix Claude Code sessions
scrubbed→
Environment tasksSFT pairsDPO preferencesDistillation set
01 · Make

Turn traces and codebases into benchmarks

Import traces from Langfuse, LangSmith, Phoenix or your coding agent, or let ArkenOS, Claude Code, Codex or Cursor build from a codebase. How making works

GSM8K200 questions · seed 0

Qwen2.5-1.5B14.0%
SmolLM3-3B9.5%
Qwen2.5-0.5B5.0%
Run #4 errorednot scored
02 · Evaluate

Any model, any benchmark

Scored by the benchmark's own verifier, on your runner. Errors are never scored. Evaluate

Each round kept only if it improvesGRPO · held-out

r1
r2
r3
r4
r5
r6

Round 4 was worse, so it was reverted0.71

03 · Train

Better every round, or reverted

Held-out tasks after every update; a worse round is rolled back. Train

Bountyscore bounty

$Beat 0.80 on fix-gitreward $500
Run submitted0.83
Target reached≥ 0.80
Accepted by posterawarded
04 · Share & earn

Publish under your name, earn through bounties

Make it public in the hub with attribution, and take on bounties: a poster names a reward for a score, a build or a fix, and accepts the winning submission. Share and earn

What we stand for

Neutral, measured, free to start.

Neutral

Bring any model, compute, sandbox or trainer and pay them directly. Nenyax orchestrates; it owns no compute.

you pay them directly

Measured

Every environment gets a tier from running it, and every score comes from its own verifier.

verifier_discriminatesaudited

Free to start

Sign up and use the platform free. Upgrade when your team needs more. See pricing

Free· Pro · Team · Enterprise
Built to plug in

Your stack, your keys, our orchestration.

Nenyax owns no compute. It runs the loop across what you already have.

01Any formatsix SDKs, six shapes02One contractManifest · Trajectory · Judgment03Same trajectorywhatever went in
Any SDK

Every format, one contract

Verifiers, OpenEnv, NeMo Gym, Harbor, Gymnasium, reasoning-gym: one call, the same trajectory. The SDK

StackChange

ModelvLLM · Qwen3-8B
ComputeRunPod · your account
SandboxModal
Trainerverl
Your stack

Swap any piece

Change the model, compute, sandbox or trainer; the environment never notices. Your stack

Your machine runner online

OPENAI_API_KEY HF_TOKEN WANDB_API_KEY
Outbound only:heartbeatclaimeventsfinish
Your machines

Keys stay home

The runner dials out from your machine. Model keys and data never leave it.

Indexed from every hublistings

Prime Intellect1,777
Hugging Face452
OpenAI Evals432
lm-eval-harness208
reasoning-gym106
The hub

3,000+ listings

Indexed from Prime Intellect, Hugging Face, OpenAI Evals and more. Explore

Build with your agentnenyax skill install

ArkenOS Claude Code Codex Cursor
TasksVerifierReward
Your agent

Build with ArkenOS or any agent

Install the skill; your coding agent writes tasks, verifier and reward. Make

Three ways in

Python, terminal or config.

The same contract underneath each one: load any environment, run any model on it, get one trajectory back.

quickstart.py
import nenyax
model = nenyax.Endpoint("http://localhost:8000/v1", "Qwen/Qwen3-8B")
env = nenyax.load("verifiers:gsm8k")
traj = env.rollout(model)
print(traj.score, len(traj.model_calls))

Load an environment from any format and roll it out with any OpenAI-compatible model.

Who it's for

Built for people who ship models.

AI engineers

Turn production traces into a benchmark for your own task, then check every model change against it.

Make a benchmark

Researchers

Evaluate models across every hub's environments with honest tiers, and train with the learner you trust.

Evaluate models

Agent builders

Build environments from your codebase with ArkenOS or any coding agent, and improve your agent in them.

Build from code

Labs and teams

Run everything on your own compute and keys, with a private hub for your organization.

See plans
Questions

Good to know.

The questions people ask before their first run.

A platform that turns your traces and codebases into benchmarks and RL environments, evaluates and trains any model on them with the stack you already use, and lets you share what you make.

Start with what you already have.

A trace export, a codebase or a model endpoint is enough to begin. Free to start.

Get started Read the docs
nenyax login
nenyax check reasoning_gym:basic_arithmetic
audited your first check