> ## Documentation Index
> Fetch the complete documentation index at: https://arkor-92aeef0e-eng-716.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Roadmap
> What we're building next in Arkor.
Arkor is in alpha, so this page is intentionally sparse. Items are grouped by what state they're in: actively being built, scoped and waiting their turn, or under consideration. We don't commit to dates yet.
## In progress
Let a fine-tuned model return typed fields (for example `{ urgency, nextAction }` from a triage run) instead of free-form text, with the output schema declared in your project as a TypeScript type.
Surface a stable HTTPS endpoint URL for each fine-tuned model so you can call it from your product without going through the SDK.
Scaffold a project without picking a starter template. Today `create-arkor` and `arkor init` require choosing `triage`, `translate`, or `redaction`; this opens the door to starting from a blank trainer.
## Up next
Silent refresh on expiry, so long-running sessions stop getting interrupted by re-login.
Upload a local JSONL file as the training dataset, alongside the existing HuggingFace name and blob URL paths.
Run training on your own GPU instead of routing every job through Arkor's managed GPUs.
Surface the existing dry-run option in the Studio UI for fast smoke tests before kicking off a full training run.
Cover the full Gemma 4 family in the `model` field. Today every starter template uses `gemma-4-E4B-it`; opening up the rest of the family lets you pick the size that fits your latency and quality budget.
## Backlog
Run the training backend on your own infrastructure, with a documented `ARKOR_CLOUD_API_URL` knob and versioned API guarantees.
Grow `createArkor` into an umbrella for shipping and evaluating models, not only training.
Export a trained model as a file you can run on your own machine or deploy target, instead of staying on Arkor's managed inference.
Generate training data from a small seed set, for cases where you don't already have a labeled dataset.
Templates that pair compatible teacher and student models so distillation runs work out of the box.
Templates aimed at small models suitable for WebGPU and mobile targets.
Expand support beyond the Gemma 4 family to additional open-weight model families, so the `model` field becomes a real menu instead of a single supported value. We're concentrating on the Gemma 4 family in this earliest stage so we can deliver fast, stable inference and training to as many people as possible.
Fine-tune on image (and eventually audio) inputs alongside text. Today every template is text-in, text-out.