Tools

Blind side-by-side comparisons voted on by the public, which produces the most honest ranking available. You can also just use it to try models for free.

Deciding between two models without reading a single benchmark table.

Independent measurement of quality against speed and price, updated as models ship.

The one place that puts cost next to capability, which is the comparison that actually matters.

One account and one API key reaching hundreds of models from many providers, with live pricing published openly.

Trying many models without signing up to many providers. Also the source of the pricing data in these appendices.

Where open models actually live. Every model card carries its licence, and most have a browsable demo.

Downloading weights, and checking what a licence really permits before you build on it.

A Chinese model hub that often has releases before they appear elsewhere, and hosts mirrors when the main hub is slow.

Getting a model whose Western download is rate-limited or unavailable.

A curated list of models packaged for one-command installation, with sensible quantisation already chosen.

The fastest route from nothing to a running model. Start here rather than at the hub.

A generous free playground for Google's models, with an API key you can use in code.

Building something real without paying anything to begin with.

Serves open-weight models on specialised hardware at unusually high speed.

Finding out how much latency actually affects your experience. Speed changes how you use a model.

Another unusually fast inference service, with a usable free tier.

The same experiment as above, on different hardware.

A self-hosted chat interface that connects to local models and hosted APIs, with documents and users.

Turning a local model into something a household or small team can actually use.

A desktop chat app for local models, open source and offline by default.

Anyone who wants a private assistant with no terminal at all.

Node-based interface for image and video generation. Steep, and nothing else comes close for control.

Serious image work, reproducible pipelines, and anything involving video.

The interface that made local image generation popular, with a large library of extensions.

Legacy workflows and old tutorials. New work is better served by ComfyUI.

Thousands of hosted demos, one per model or tool, runnable in a browser.

Seeing whether a model is any good before downloading twelve gigabytes of it.

Notebooks with a GPU attached, free at a basic tier.

Running something once, without owning hardware or configuring a server.

Notebooks with a weekly GPU allowance, and a large public dataset collection.

Free GPU hours, and data to experiment with.

Rent a GPU by the hour, with templates for common inference setups.

Running a model too large for your machine, occasionally. Remember to stop the instance.

The most widely used framework for connecting models to tools, data and each other.

Prototypes, and anything you might hand to someone else. Ubiquity is its main virtue.

A framework focused specifically on retrieval over your own documents.

Private document search — the use case most people actually want from local AI.

A TypeScript toolkit for streaming model output into a web application.

Putting a model behind a web interface without reinventing streaming.