Leading Open-Weight Models: A More Cost-Effective Path to Sovereign AI in Europe

For a long time, nearly every AI procurement decision came down to the same trade-off: organizations that wanted the most capable models had to accept US providers, US infrastructure, and US contractual terms. Organizations that wanted to remain sovereign in Europe had to accept lower performance. Over recent months, though, that trade-off has become obsolete.

Effective immediately, our customers can choose from three of the strongest Open-Weight models available: Kimi K3, DeepSeek V4, and GLM 5.2. All three run entirely on European infrastructure. This enables a productive AI environment in which no US provider is involved, without compromising on top-tier performance.

Open-Weight Models Lead on Price-Performance

In the GPQA Diamond benchmark, widely considered the industry's most demanding reasoning benchmark, Kimi K3 scores 93.5%, making it the strongest Open-Weight model available. Grok 4.6 leads the overall ranking at 94.9%. The gap between the best open and best closed model is therefore just over 1.5 percentage points. Eighteen months ago, the gap was still in double digits. The remaining models follow closely: DeepSeek V4 Pro reaches 90.1%, while GLM 5.2 scores 89%.

The LLM Stats leaderboard also shows how capable these models have become. Four of the current top ten models are Open-Weight. LLM Stats aggregates results across several benchmarks into a single score. The result is clear: Open-Weight models now solve the same kinds of tasks as the frontier models offered by major US providers, while in some cases costing substantially less, as the chart below shows.

From this, a clear picture emerges: organizations that currently rely primarily on leading US models can get significantly further with the same AI budget by using these new options.

Open Weights Enable Genuine Sovereignty

An Open-Weight model is an LLM whose trained parameters, referred to as weights, are publicly available. It can be downloaded, operated independently, and in many cases adapted. A proprietary model can only run in Europe if its provider chooses to make it available there. In contrast, anyone with access to the weights and the necessary infrastructure can operate an Open-Weight model.

Sovereignty is therefore no longer merely a provider's promise but ensured through the licence under which the model is released. nuwacom works with the German inference API provider Melious who exclusively uses models running on infrastructure inside the EU, operated by local companies.

nuwacom Provides European-Sovereign AI Infrastructure for Enterprises

The new Open-Weight models add another building block to our sovereignty architecture. They create the option to build an AI operating system that works entirely without US models while still meeting the highest performance requirements.

Other important components of this sovereign architecture include our partnership with STACKIT, which enables customers to operate their entire AI infrastructure in Germany, and PII scrubbing, which anonymises sensitive information in the input before it is sent to a model.

Combined with the three new models, this creates a possible configuration in which sensitive data never leaves the EU and no US provider is involved at any point. It gives organizations with demanding data protection and compliance requirements a high-performance option for using AI productively and sovereignly.

Which Model for Which Task

DeepSeek V4 is well suited to high-volume routine tasks: classification, extraction, summarizing, and standard correspondence. For these workloads, cost per task matters, while the difference in output quality, compared to the strongest available model, is often barely noticeable.

GLM 5.2 is the natural choice when throughput and quality matter equally. It is a strong option for everyday chat tasks, document analysis, and writing shorter texts.

Kimi K3 is most useful where a highly capable model is needed: complex tasks, creating and analyzing large documents, and agentic work.

All three models are now available in every nuwacom workspace. Our recommendation: choose a real task from your daily work, have your current model and the new models complete it, then compare the results. That is the most meaningful benchmark - and the one most relevant to you.

Try nuwacom free for seven days or book a demo with one of our AI experts.