Build and run your own models in TypeScript.

01Define

src/arkor/index.ts

import { createArkor, createTrainer } from "arkor";

const trainer = createTrainer({
  name: "support-triage",
  model: "unsloth/gemma-4-E4B-it",
  dataset: { ... },
  lora: { r: 16, alpha: 16 },
  callbacks: { onLog, onCheckpoint },
});

export const arkor = createArkor({ trainer });

02Train

arkor start

$arkor start

→ step 10 · loss 1.42

→ step 50 · loss 0.74

→ step 100 · loss 0.41

→ step 200 · loss 0.28

✓training complete

✓adapter saved

03Test mid-training

src/arkor/trainer.ts

Live

onCheckpoint: async ({ infer }) => {
  const res = await infer({
    messages: [{ role: "user", content: ticket }],
  });

  // → category: billing
  // → urgency:  high
  // → action:   forward
}

Arkor is a TypeScript framework for fine-tuning open-weight models on managed GPUs, served behind a hosted OpenAI-compatible endpoint. Designed for product engineers shipping LLM features. No signup to try locally. Works with any Node app — Next.js, Hono, Express, anything.

Owning the model should feel
like shipping software.#

API calls don’t give you ownership

The feature works, but the model is not yours. You can't tune it for your data, run it where you want, or control the margin when pricing changes.

ML infrastructure is built for ML teams

GPUs, training jobs, adapters, and inference stacks usually belong to platform teams. Product engineers should not need that stack just to improve app behavior.

Prompting and RAG only go so far

Some behavior needs to live in the model itself: consistency, latency, and domain-specific decisions. At that point, you change the model, not the prompt around it.

Pick your path.#

I want to use an open-weight model in my app.

No training. Just a hosted Gemma 4 endpoint.

App Server

I have a dataset and want to fine-tune.

Code the run in TypeScript, train on managed GPUs, inspect via Studio.

FrameworkCloudStudio

I want to optimize the model my app uses over time.

Start with a hosted model. Distill it down to a smaller, cheaper, faster one as you learn what you need.

App ServerDistil

Three beliefs that shape every product.#

Small models perform well on backend semantic tasks.

See it yourself

≈

31b

Open-weight means you keep what you fine-tune, run it anywhere, and avoid API lock-in.

Keep what you fine-tune

Your adapter is yours.

Run anywhere

Self-host, on-prem, or edge.

No API lock-in

No closed model dependency.

TypeScript means it fits where your app already lives.

Product logicBackendsInternal toolsCI / deployIntegrations

Live demo

See it in the playground.#

Run a support-triage call against a base open-weight model, a fine-tuned version, and a frontier reference. Same input, same prompt.

Support triage

A customer message arrives. The model reads intent, assigns a category and urgency level, and recommends a next action. No keyword rules, no routing trees.

Base Model

gemma-3-4b

Waiting for input.

Fine-tuned

gemma-3-4b (fine-tuned)

Waiting for input.

Reference

gemini-3-flash-preview

Waiting for input.

See all three demos

Your model.
Your endpoint. Your terms.#

Start for free with no signup, or set up a hosted endpoint in under a minute.

Start for free

Build and run your own models in TypeScript.

Owning the model should feellike shipping software.#

API calls don’t give you ownership

ML infrastructure is built for ML teams

Prompting and RAG only go so far

Pick your path.#

I want to use an open-weight model in my app.

I have a dataset and want to fine-tune.

I want to optimize the model my app uses over time.

Three beliefs that shape every product.#

Small models perform well on backend semantic tasks.

Open-weight means you keep what you fine-tune, run it anywhere, and avoid API lock-in.

TypeScript means it fits where your app already lives.

See it in the playground.#

Support triage

Your model.Your endpoint. Your terms.#

Owning the model should feel
like shipping software.#

Your model.
Your endpoint. Your terms.#