Home›Products›Model Catalog
Nobody wants an empty GPU; they want something that runs. A template is one image plus sane defaults — disk, ports, a recommended GPU. The "pick one, start" flow on the home page and the console's rent dialog are fed by this same list.
Console links point to the console domain; when you open the site locally they are rewritten to localhost:3000.
Where things stand
Our status rule is one sentence: a template gets called ready when it runs today from a public, official image, and nothing else gets called ready. We would rather keep the list short than put a "ready" label on something that is not.
A green dot here means "written as part of Phase 0 and running in local development" — it is not a statement about production health. Component by component: Status.
Starting templates
There is no price column: the price follows the GPU you pick, not the template. The recommended GPU below is a sensible starting card for that kind of work; current hourly rates are on the pricing page.
| Template | What it does | Category | Recommended GPU | Min VRAM | Status | Start |
|---|---|---|---|---|---|---|
| Jupyter + PyTorch | Experiment and prototype; from a notebook to a training script | dev | RTX 4090 | 24 GB | ready | Start → |
| vLLM · OpenAI-compatible server | Publish a model as an API; only base_url changes on the client | inference | A100 80GB | 48 GB | ready | Start → |
| Ollama · your own chat | Pull and run open models with a single command | inference | RTX 4090 | 24 GB | ready | Start → |
| Axolotl · fine-tuning | LoRA/QLoRA and full fine-tune | training | A100 80GB | 48 GB | ready | Start → |
| ComfyUI · image generation | SD/SDXL/Flux workflows, node-based interface | image | RTX 4090 | 24 GB | preparing | Start → |
| Whisper · speech transcription | Transcription, subtitles, batch audio processing | audio | RTX 3090 | 16 GB | preparing | Start → |
| Blender · render | 3D render, animation, architectural visualization | render | RTX 4090 | 24 GB | preparing | Start → |
| Agent runtime | Tool-calling LLM agents, long-running jobs | agent | RTX 4090 | 24 GB | preparing | Start → |
| Custom image | Your own container — paste the registry address | dev | — | — | ready | Start → |
preparing means: the template package is still being put together and the image is not pinned yet. You can run that work today anyway — "Start" hands you the custom-image flow, and the only difference is that you supply the image. We deliberately do not print image tags on this page: tags move, pages go stale, and the current image sits in the console's rent dialog.
Categories
A category is not just a label: the defaults behind it decide which ports the pod opens, how much disk it asks for and which card is a sensible starting point. Where a template is still preparing, the work itself is possible today anyway — with your own image.
Jupyter + PyTorch for the notebook-to-script path, and the custom image for everything the catalog does not cover. Both run today.
vLLM turns an open model into an OpenAI-compatible API; Ollama is the quickest path to your own chat. Detail: Inference.
Axolotl covers LoRA, QLoRA and full fine-tunes; checkpoints belong on a volume that outlives the pod. Detail: Fine-tuning.
ComfyUI's node graph for SD/SDXL/Flux. The template is preparing; today the same workflow runs from your own image. Detail: Image & video.
Whisper for transcription, subtitles and batch audio jobs. The template is preparing; today it runs from your own image, and a 16 GB card is enough.
Blender for 3D render, animation and architectural visualization. The template is preparing; today it runs from your own image.
An agent runtime for LLMs that call tools and jobs that stay up for hours. The template is preparing; today it runs from your own image.
Multi-node cluster training belongs to Clusters — after Phase 2, soon. A scale-to-zero production endpoint belongs to Serverless Inference — Phase 2, soon. Neither is a template, and neither exists today.
What a template is
A template does not put a layer between you and the GPU. It answers the three questions you would otherwise answer by hand before the first log line.
Every template carries a volume size, the ports it needs and the card it is comfortable on. You can override all of them in the rent dialog; the point is that the defaults are not a guess you have to make at midnight.
Each Start above opens the console's rent dialog with that template preselected. If a link ever carries an id we no longer recognise, it is ignored quietly and the screen still opens — a stale bookmark should not break a page.
The minimum VRAM figures are what the template comfortably fits into. The real choice follows your model size and context length; the marketplace will happily rent you something larger, and the pricing page shows what that costs. On a stopped pod, storage still bills — that line is visible on the billing screen.
Ready-made templates are a convenience. Verifying that a catalog image is the one we published, and having it already sitting on the host, is the part that touches security and money. Both are Phase 2 work, and neither is running today.
Pre-pull is a price feature, not a speed feature: pulling ten gigabytes while the pod boots disappears into the noise of an hourly bill, but on a per-second bill it comes straight out of your pocket. Pre-pull and the model cache both need persistent space on the host side, which depends on our own host network — Phase 1 and Phase 2 in the technical plan. The same mechanism is what would make the cold-start target of Serverless Inference possible.
From code
The same list feeds this page, the console's rent dialog and your own scripts. If you would rather stay in the terminal, none of this needs a browser.
Contribute
Which four get packaged next is decided by the work people actually run. Tell us what you run, and the list gets better rather than longer.
The mailbox is a placeholder and opens soon.