Every model here is one key, one bill, one European counterparty.

How to choose a model

Start from the task, not from a leaderboard. Each group below links to the model pages; the figures behind them come from the developers' own model cards.

Which languages do the developers name?

A model can work in a language its developer does not list — but only a listed language is a documented claim. This is what each model card says; where it says nothing, test with your own material.

Open weights and licences

“Open weights” means the trained parameters are public; the training data usually is not. What you may do with a model is decided by its licence — through our API as much as on your own server.

Through an API or on your own server?

Through the API you download nothing: the model is already running, you send an OpenAI-compatible request with base_url https://api.heabsy.com/v1 and the model ID, and you pay per token. Switching models means changing one string.

Running weights yourself makes sense when a model is small and the load is steady — OpenAI states that gpt-oss-20b fits into 16 GB of memory. The largest open models in this catalogue are a different matter: Kimi K3 has 2.8 trillion parameters, which puts it at data-centre scale. Claude and the GPT-6 models cannot be run yourself at all — their weights are closed.

Terms used in the catalogue

Token

The unit a model reads and writes — roughly a short word or part of a word. Prices are per million tokens, separately for input and output.

Context window

How many tokens fit into one request, including the conversation so far and the documents you attach.

Cached input

Input the model has already processed in an earlier request of the same conversation. It is billed at a lower rate — in agent work it is usually the largest share.

Reasoning (thinking)

The model writes intermediate reasoning before the answer. It improves hard tasks and costs output tokens and time.

Mixture of experts, active parameters

Only part of the network runs for each token. A model with 284B parameters and 13B active is as large as the first number and roughly as fast as the second.

Zero data retention (ZDR)

Prompts and completions are not stored by whoever runs the model. Our own models run this way; partner-hosted models do not come with that guarantee — each model page says which applies.

Questions about the catalogue