Open-Source LLMs: Transparency, Control and Protection of Your Investment

«For the most part don't really work – and are not smart.»

Satya Nadella, the CEO of Microsoft, said this internally about Microsoft’s own Copilot integrations. Source: The Information, 28 December 2025.

Not a critic. Not a competitor. The CEO of the company that has invested billions in OpenAI and AI since 2023.

That is the starting point for this post.

What Open Source and Open Weight mean for LLMs

In the software world, open source means the source code is public, can be modified, and can be redistributed. For AI models, it is more complex. The terms are often confused.

Open Source in the strictest sense of an LLM means the training code, training data and model weights are fully disclosed. Real open-source models in this sense are rare.

Open Weight means the trained weights — in simple terms, the model’s "brain" — are publicly accessible and can run locally. The training process itself does not need to be fully documented. Most models that are loosely called "open source" are actually open weight.

Well-known open-weight models: Llama (Meta), Gemma (Google), Phi (Microsoft), Mistral (in part) and Apertus (a Swiss initiative). Tools such as Ollama or LM Studio install them directly on your own computer or server — no cloud upload, no subscription.

The key difference from proprietary models such as GPT-5.5, Claude or Gemini: you own the model.

The black box problem

With closed models from OpenAI, Anthropic or Google, you do not know:

• What it was trained on. Copyright issues are unresolved. Several lawsuits are under way in the US. Providers disclose their training data selectively.

• How it is weighted. Every model has a system prompt and internal optimisations. It is not transparent who benefits from them.

• What happens to the data you enter. Pro licences exclude training on your chats — that is in the terms and conditions. Whether this is really handled that way cannot be checked from the outside.

• What the US Cloud Act means. US authorities can access servers of US companies if they have a justified suspicion — even if the servers are in Europe. For Anthropic, OpenAI and Google, this is legally relevant.

In short: you send data into a black box and have to trust that nothing happens to it. That trust is not irrational — but it is still only trust.

Open Weight in practice: Apertus as a Swiss example

Apertus is a Swiss LLM initiative with an approach that is unusual for the industry: the model was deliberately trained only on copyright-safe material. News articles were left out — not because of quality problems, but because tests showed the model performed just as well without them. That is a decision you can document.

That is the difference from a black box: you can see what is in the model and what is not.

Version 1 reached Llama level. Version 1.5 (8B parameters) arrives in May 2026.

If you run Apertus on a Swiss server or internally, you know exactly this: the data does not leave the defined perimeter. No Cloud Act. No terms-and-conditions clause. No dependence on the goodwill of a US corporation.

Protecting your investment: three scenarios you should plan for

Geopolitics. In a trade conflict, the US government can decide that certain industries or countries no longer have access to US AI services. That sounds abstract — but it is legally possible and has already been applied in other technology areas, such as chips and software. A locally run open-weight model is not affected by this.

Vendor instability. The valuations of the major AI providers are historically high. How long investors will keep funding ongoing losses is uncertain — especially for companies such as OpenAI, whose financing structure and USD 340 billion valuation are based on very optimistic growth assumptions. Google and Anthropic are more sustainably positioned — but even with cloud providers, the rule holds: products are retired, prices change, APIs are deprecated. Anyone who builds an AI workflow on an external service has to keep adapting.

Vendor lock-in. Every customisation, every integration, every workflow you build on a proprietary model depends on that model surviving. A locally run open-weight model can be replaced whenever you want — with no data migration and no contract termination.

Copilot: a special case with double dependency

Microsoft Copilot has two structural problems that other cloud models do not have.

First: double dependency. Copilot is a Microsoft layer built on OpenAI models. You depend on Microsoft and OpenAI at the same time — with all the risks of both providers.

Second: system constraints. Copilot is not the same as direct access to GPT-5.5. Copilot routes complex requests to GPT-5.5, simple requests to GPT-5.3 Instant or GPT-4o — Microsoft chooses the model, not the user. System prompts, guardrails and the integration into Microsoft 365 also change model behaviour. Anyone using ChatGPT directly usually has more room to manoeuvre and better results on the same tasks.

The numbers confirm it:

3.3% of the roughly 450 million Microsoft 365 customers pay for Copilot. That is about 15 million seats — after billions invested in marketing and integration. (Source: Piper Sandler / myBusinessFuture, February 2026)

5% of companies that start a Copilot pilot move on to a larger deployment. 95% get stuck in the pilot phase because the business case or data quality is not good enough. (Source: Gartner, July 2024)

–24 is Copilot’s Accuracy Net Promoter Score. Users trust the answers less than they trust them. A negative NPS means more users actively recommend against the product than recommend it. (Source: Recon Analytics, September 2024)

Satya Nadella called it internally “for the most part don't really work – and are not smart”. The numbers back him up.

Conclusion: what you are really weighing up when choosing a model

Proprietary cloud models such as Claude, ChatGPT or Gemini are powerful, quick to deploy, and actively developed. For many use cases, they are the right choice.

But if you work with sensitive data, operate in regulated industries, or want to build on AI over the long term, you should ask three questions:

1. Control: Do I know what happens to my data?

2. Traceability: Can I understand and check the model’s behaviour?

3. Investment protection: Am I dependent on a provider whose future I cannot influence?

Open-weight models give a clearer answer to all three questions than any proprietary system.

That does not automatically make them the better choice — but it does make them one you should know about.

What comes next

Now the basics are in place: which LLM, which model, which licence — and what control over the model really means. In the next post, we will look at how you connect Claude to your own tools: connectors, MCP, and how a wording skill is created that transfers your writing style into every output.

→ Connectors, MCP and your own wording skill