Open-weight AI models take the lead: what it means for nonprofits

Artificial intelligence has long been something you rent from a few large providers. That is changing. Open-weight models – where the weights are freely available – now match the next-strongest proprietary models on real tasks, at a fraction of the cost. For small organisations and nonprofits, that is an opening worth understanding: both the possibilities and the limits.

What happened

The Chinese model GLM-5.2 (from Z.ai / Zhipu AI, released 13 June 2026) is a clear example. It has 744 billion parameters in total and 40 billion active (a mixture-of-experts design), is released under an MIT licence with the weights freely on Hugging Face, and is served at roughly a sixth of the price of the proprietary flagships. On well-regarded coding benchmarks it beats GPT-5.5: on SWE-bench Pro, GLM-5.2 scores around 62% against GPT-5.5’s just over 58%, and it also leads on FrontierSWE.

At the same time, the centre of gravity in usage has shifted. According to a recent study, Chinese open-weight models now carry around 60% of token traffic on neutral routers such as OpenRouter – and the share is rising. A year ago, proprietary models dominated. Open models are no longer a niche choice.

An important nuance: “matches” is not “best in the world”

It is easy to oversell this, so let us be precise. GLM-5.2 matches and beats GPT-5.5 on several coding tasks – but the very strongest proprietary models still lead. Claude Opus 4.8, for instance, remains ahead of GLM-5.2 on the same benchmarks (around 69% against 62% on SWE-bench Pro, and 85 against 81 on Terminal-Bench). The point is not that open models have taken first place, but that the gap to the top has become small – and that you get a great deal for zero licence cost.

Open weights are not the same as open source

Here is a distinction worth understanding. That the weights are open means you can download the model, run it and fine-tune it on your own data. It does not necessarily mean the training data and the full recipe are public – as fully open source would require. The difference determines how free you actually are, and how well a community can scrutinise the model. An MIT licence on the weights is nonetheless very permissive compared with closed alternatives, and gives you rights you never get from a pure rental service. The idea is the same as behind free software for nonprofits: a digital commons that many can build on.

What it means for small organisations

For a nonprofit or a founder on a tight budget, this is more than a technical curiosity:

• Real access without licence costs. You can use capable AI without an expensive per-user subscription. That levels the field between large and small players.
• Privacy and control. Open weights can be run locally or with a provider you choose yourself. Then sensitive data need not leave your machine unnecessarily – useful for organisations handling member or client information.
• Independence. You are not tied to one provider’s prices, terms or shutdowns. The model you have downloaded does not vanish if a company changes strategy.

Concrete uses

You do not need to be a technologist to benefit from this. Typical tasks where an open model can save time in a small organisation:

• Drafting applications, minutes and information texts – which you then quality-check yourself.
• Translating between Norwegian and English, for example for web pages or newsletters.
• Summarising long documents and meeting notes into short points.
• Answering frequently asked questions based on the organisation’s own documents.

How to choose a model and provider

Few small organisations will run a large model themselves – it requires hardware, setup and maintenance. The realistic choice is often a hosting service that offers the open model: you avoid the operations, but keep the freedom of choice and can switch provider later. If you handle personal data, look for a provider in Europe with a data processing agreement, so that you meet the privacy rules. If you want to run entirely locally for privacy reasons, there are smaller open models that run on ordinary office hardware – weaker than GLM-5.2, but good enough for many tasks.

Be honest about the limits

Open models do not solve everything. They can be wrong, invent facts and reflect biases in their training data, just like proprietary models. The responsibility for what you publish still rests with you. Whichever model you choose, the same ground rules for responsible use apply – we have gathered them in responsible AI for nonprofits.

A more level playing field

For a nonprofit, the biggest thing about this shift may not be the technology itself, but what it does to access. When capable AI existed only behind expensive subscriptions, it was the large and well-funded who could put it to use. When the weights are open and the price a fraction, a small association, a local chapter or a founder with an empty wallet gets much of the same power. That is a levelling close to our own purpose: that digital tools should be available to everyone, not only to those who can afford them.

At the same time, responsibility comes with access. Use it to free up time for what the organisation actually exists to do – not to replace the human judgement the work requires.