How do Apertus, OLMo, SEA-LION and BLOOM compare?

"Open" and "public" get used as if they meant the same thing. They do not. A model can be open but privately governed, or public but built on a company's foundation. Comparing these four projects is the quickest way to see the difference, and to choose one for a real job.

Three separate questions sort them out. Who controls it? That means a public institution, a nonprofit or a volunteer collaboration. What is published? That can be only the final weights, or also the data and training code. What may you do with it? That depends on the licence.

The four side by side

"B" means billion. Sizes describe different kinds of models, so they are not a score. The 30B and 120B SEA-LION models are mixture-of-experts, which means only part of the model runs for each token (3B and 12B active, according to their names). A bigger number therefore does not mean a slower or better one.

What each one is good for

Apertus is the clearest case of public AI: public institutions, public compute, and what the team says is full transparency. The 1.5 update, according to the Public AI announcement, added images, optional reasoning, a context window of 262,144 tokens and better tool use. That announcement also promises a technical report later, so independent benchmark results for 1.5 are still thin. Bruce Schneier and Nathan Sanders, who favour public AI, call it the world's most advanced public model and also note that it is about the size of the 2023 generation of commercial models.

OLMo is the researcher's choice. Ai2 calls its approach the "model flow": the whole path from data to final weights is published, so you can trace a behaviour back to the training decisions behind it. Olmo 3 comes in 7B and 32B sizes that Ai2 says run on anything from laptops to research clusters, and in March Ai2 released Olmo Hybrid, a 7B experiment mixing transformer attention with recurrent layers. Ai2 is not a government body, however. Schneier and Sanders point out that it relies on corporate partners for compute and is not under public control. And while the licence is Apache 2.0, the model card describes the models as intended for research and education.

SEA-LION solves a practical problem that large labs underserve: Southeast Asian languages such as Burmese, Filipino, Thai, Tamil and Vietnamese. Its strategy has shifted. Earlier versions were trained from scratch, while newer ones start from someone else's model. Version 4.8 was built with NVIDIA on its Nemotron-3 bases. That brings strength for less money, but you inherit choices you cannot fully inspect. Schneier and Sanders describe this as a hard trade-off between cost and usefulness, not as a failure.

BLOOM matters mostly as history and as proof of concept. In 2022, volunteers working with the French government's Jean Zay supercomputer showed that a large open model could be trained outside a big company. Its RAIL licence is the main thing to note: it restricts certain uses, so it is not open source in the way Apache and MIT are. At 176B and with a 2022 design, it needs a lot of memory for its age, so it is better treated as something to study than as a first choice to deploy.

Others are worth watching. Stanford's Marin describes itself as training a 535B mixture-of-experts model on 18 trillion tokens. Spain's ALIA and the European EuroLLM are also in this space, and the Public AI site says its network helps host them.

A worked example: four projects, four picks

These scenarios are illustrative.

For a commercial product, read the licence file yourself. Apache and MIT are short and permissive. RAIL has conditions. And a permissive licence on a model built from another model still deserves a check of the parent's terms, since SEA-LION's parent is an NVIDIA base model with its own licence.

How much hardware do they need?

A rough rule: each parameter takes about 2 bytes at 16-bit precision, and about half a byte when quantized to 4 bits, a compression that trades some quality for size.

These figures are arithmetic and leave out extra working memory. Some models and versions also require specific software support, so a quantized file may not exist for each one yet. Check the model page first. Our VRAM calculator goes further.

Which one should you try first?

If you only want to see what public AI feels like, start with Apertus at the Public AI chat, where the 8B reasoning version needs no login. It takes a minute, and it gives you a real sense of where these models stand.

If you want to run something on your own machine, try a 7B or 8B model first, from either Apertus or OLMo, since they fit on an ordinary laptop when quantized. Move up only when you hit a limit you can name. If your work is in a Southeast Asian language, go straight to SEA-LION, since neither of the others is tuned for it.

Be realistic about what you are testing. None of these aims to match the strongest commercial or open-weight models. Our hub on open-weight models covers those. You would choose one of these four because you value what they let you see, and which languages they serve, as much as raw ability.