Kolibri-1 Debuts Open Weights as Google Cuts Gemini

A German 78-billion-parameter MoE model targets local data control, while the search giant restricts unpaid users to trim server bills.

Two divergent strategies in artificial intelligence collided in early October 2026. German developer Aleph Alpha published the model weights for Kolibri-1 under an Apache 2.0 license on Hugging Face, designed for organizations that cannot send private data to external clouds. In the same week, Google moved in the opposite direction for public access: starting October 9, 2026, unpaid Gemini accounts lose larger models and drop down to Gemini Flash-Lite.

How Kolibri-1 operates

Published on October 3 to mark German Unity Day, Kolibri-1 runs on a Mixture-of-Experts (MoE) architecture. Rather than firing every parameter for every word, the network routes queries through smaller specialized sub-networks called experts. Kolibri-1 packs 78.1 billion total parameters, but activates only 3.46 billion per token during inference. That selective routing cuts memory consumption enough to let teams run the system on local clusters using serving engines like vLLM.

The context window reaches up to 1,048,576 tokens, covering hundreds of pages of technical documentation. To preserve responsive latency in production, Aleph Alpha recommends keeping active inputs below 262,000 tokens. The pre-training corpus included 21.3 percent organic German data, tailored to satisfy the European Union AI Act and GDPR compliance standards.

The release, however, is not a full open-source package. The Apache 2.0 license covers only the numerical weights generated during training. The proprietary training codebase, raw datasets, and alignment pipelines remain private trade secrets. It is an open-weight release rather than open source. For European hospitals, banks, and government agencies, that separation is sufficient: compute stays inside company walls, avoiding external data transit.

Access cuts across Google Gemini tiers

While Aleph Alpha pushes execution onto private hardware, Google restructured its consumer lineup on the web and mobile apps to curb spiraling infrastructure overhead. Company support pages confirm that free accounts lose both standard Gemini Flash and Gemini Pro on October 9. Personal users without paid subscriptions will interact exclusively with Gemini Flash-Lite, the lightest tier in the family.

Entry-level paying subscribers also face downgrades. The Google AI Plus plan at $4.99 per month loses Gemini Pro, keeping only Flash and Flash-Lite. Access to Gemini Pro now requires the top-tier AI Pro plan at $19.99 monthly. That premium subscription adds Deep Think, an extended reasoning tool with three selectable compute levels: low, medium, and high. Google noted that commercial Google Workspace accounts operate under existing enterprise contracts and are exempt from these consumer changes.

The practical cost of compute

Both announcements reflect the same economic reality handled through contrasting playbooks. Cloud providers have realized that subsidizing heavy reasoning for hundreds of millions of casual users damages margins. Google’s response narrows free utility, herding power users into recurring $20 subscriptions.

Kolibri-1 takes the inverse path by marketing data sovereignty. Downloading the weights removes dependencies on foreign cloud APIs and eases regulatory scrutiny. The trade-off is capital expenditure: the cost of silicon, cooling, power, and operational staff moves off cloud bills and lands squarely on the enterprise’s own balance sheet.