Open Weights in October 2026: Why open models are enough for almost everything

As the operator of an AI platform, we see every day what our users actually work with. And the picture is clear: around 97 percent of their work on Kasimir runs on open models. The big closed frontier models sit in the same model menu, one click away. They are rarely needed.

This isn't something we prescribe. It's a choice our users make themselves: emails, summaries, research, working with their own documents and data — for everyday work in a company, open models now deliver top-tier results. And things are only picking up speed: almost every month, a new open model arrives that can do more than the last one.

Two years ago, we still had to explain why an open model is enough for everyday work. Today the question is more what you would even need anything else for.

In this post, we look at what has happened with open models since the summer, what we pay attention to when testing them, and which pitfalls hide in their licenses.

A year in which open models caught up

2026 is the year many big model families put their open models under real open-source licenses. Google released Gemma 4 under Apache 2.0 for the first time in April, Mistral publishes its entire Mistral 3 family under that license, and so does OpenAI with its gpt-oss models. Since then, one release has followed the next.

For our customers, that's good news. The models they already get along well with keep getting stronger successors — and because Kasimir runs open models on its own hardware in Germany, we can add new models as soon as they have convinced us.

Two classes, one misunderstanding

Open models have split into two classes, and many comparison articles mix them up.

The first class holds the frontier models with hundreds of billions to trillions of parameters. Their weights are publicly available for download. Yet hardly anyone outside a large data center can run them. How stark that is shows in one sentence from DeepSeek's announcement of V4.1-Flash in September:

“Planning a large-scale deployment with 2,000 GPUs + a storage cluster? Let's talk.”

The second class holds models that fit on a single professional graphics card. That includes compact models with roughly 20 to 35 billion parameters, but also larger ones like gpt-oss-120b: OpenAI ships it already compressed, and only a small share of its 117 billion parameters is active per word. Models like these run at predictable cost and can be operated in many German data centers. This is exactly the class that has become so good over the past twelve months that it covers almost all everyday work.

How we test new models

Before a new model reaches our customers, it goes through a test phase with us. A good spot on a leaderboard isn't enough. We look at what matters in everyday work:

German: Does the model write clean, natural German, including technical language?

Tools: Can it reliably search the web, read files and work with connected services without falling out of its role?

Documents: Can it handle long documents, tables and images?

License and operation: Is the license clean, and does the model run economically on hardware in Germany?

Our self-hosted AI model has earned its place. We don’t run a model just because it currently tops a leaderboard. It has to show us that it writes clean German, handles tools reliably, copes with long documents and fits our infrastructure.

What's currently on the test bench

Qwen3.8-27B. Apache 2.0, 27 billion parameters with image understanding, very long context. Same class as the models our users work with today, which makes it a direct candidate for the next step.

gpt-oss-120b and gpt-oss-20b. Apache 2.0. In the large model, only a small share of the parameters is active per word, and despite its size it fits on a single card. That promises fast answers with broad knowledge.

Mistral Small 4. Apache 2.0, from the EU. Despite the name, at 119 billion parameters it's no longer a small model, but it only uses a fraction of them per word.

DeepSeek-V4.1-Flash. MIT license, context of up to one million tokens. Not a candidate for a single card, but interesting for tasks with very long documents.

And the next ones are already on the horizon: in early October, Mistral presented Large 4 and announced its open weights for the end of October. Alibaba has announced a new generation with Qwen4, which is again set to include a model in the 27-billion class.

The license is the part nobody reads

Not every open model is equally open. In August, Alibaba released both ends of its Qwen 3.8 series, one to two days apart and under completely different legal terms. The small model is under Apache 2.0. The large one, Qwen3.8-Max, is under its own license with a revenue threshold:

“If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, and the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000 (or the equivalent amount in any other currencies) during any consecutive twelve (12) months, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose.”

In plain terms: anyone making more than 50 million US dollars a year with model hosting or an AI work assistant needs a separate agreement. Purely internal use is exempt, as long as neither the model nor its outputs are made available to third parties. Above 100 million monthly active users or 20 million US dollars in monthly revenue, the same license also requires the model name to be displayed prominently.

Meta's Llama 4 models are under their own community license, which the Open Source Initiative explicitly does not recognize as open source. It includes a threshold of 700 million monthly active users and, for the multimodal models, excludes users based in the EU.

European models: an honest assessment

Open models also come from Europe: EuroLLM with 22 billion parameters under Apache 2.0, the Mistral family from France and Apertus from Switzerland, released in version 1.5 in July. Teuken-7B from the OpenGPT-X project was trained on all 24 official EU languages. Here it pays to check the license: the current version is under a non-commercial license, only an older version is under Apache 2.0.

What matters more than a model's origin, though, is where it runs. A model is a file of numbers. Where that file is executed and who has access to it decides data protection and data sovereignty, not the place where it was trained. We welcome every strong European model and add each one to our tests.

What this means legally

Since 2 August 2026, the EU AI Office can enforce the obligations for providers of general-purpose AI models with fines. In Germany, the KI-Marktüberwachungs- und Innovationsförderungsgesetz (KI-MIG, the national AI market surveillance act) has been in force since the end of July, with the Federal Network Agency as the supervisory authority. For models under a free license, the AI Act provides relief: a copyright policy and a summary of the training data remain mandatory, while some documentation obligations fall away. This relief doesn't apply to models with systemic risk.

What matters for most companies: anyone using a finished model is, as a rule, a deployer and not a provider. That means considerably leaner obligations than for whoever publishes the model. It doesn't replace a legal review of the individual case.

What we take away

Kasimir is built exactly for this: open models, run on hardware in Germany, without your inputs flowing into anyone else's training. Which models lead the way is allowed to change — that's exactly why we test.