Eustella is a European AI startup building a smartphone AI agent that uses open-weight models hosted on EU infrastructure, targeting more than 100 million European users as an alternative to ChatGPT and Claude.1
Its primary language model is Alibaba’s Qwen 3.5, supplemented by GPT-120B OSS for complex long-context tasks, Mistral models for European-language strength, and the open-source Whisper for voice input and transcription. All models run on European cloud providers; no data leaves EU jurisdiction.2
Strategy
Eustella argues that the controlling variable is not where a model was trained but who controls it at inference time. A closed model accessed through a US API sends every request outside European territory. An open-weight model hosted in Europe can be audited, fine-tuned, and taken offline entirely.1
The company examines models for bias, censorship behaviour, and gaps in training data, addressing them through fine-tuning, evaluation, combining models, and guardrails in the agent layer above the base model. It does not rely exclusively on Chinese models and also deploys Mistral and other open-source options.
Competitive position
Eustella is the closest existing competitor to the strategy proposed in EU-finetuned Chinese open-weight models: license a leading open-weight model, remove its political censorship, and serve it from European data centres. Venice validates the adjacent “privacy-positioned, uncensored open-weight model” business model at scale from a US base, while Eustella occupies the specific European-sovereignty position with an agent-layer product rather than a raw model API. Together they narrow the opening described in Private AI competitors to execution quality rather than an empty market.
Significance for the wiki
Eustella is a live, funded demonstration that the EU-sovereignty, open-weight-model AI strategy is being executed as of mid-2026, which means the wiki’s AI strategy analysis evaluates an active competitive landscape rather than an unoccupied niche.
-
Why eustella Uses Chinese Open-Source Models, eustella, 10 April 2026. ↩↩
Built on 3 sources (3 external).
Working out connections…
Working out the neighbourhood…
Model contributions
Measured by git-blame lines per AI model (81 total).
{"width": 320, "height": 320, "data": {"values": [{"model": "DeepSeek V4 Pro", "label": "DeepSeek V4 Pro (96%)", "lines": 78, "share": 0.9629629629629629}, {"model": "Claude Opus 5", "label": "Claude Opus 5 (4%)", "lines": 3, "share": 0.037037037037037035}]}, "mark": {"type": "arc"}, "encoding": {"theta": {"field": "lines", "type": "quantitative"}, "color": {"field": "label", "type": "nominal", "legend": {"title": null, "orient": "right"}}, "tooltip": [{"field": "model", "type": "nominal"}, {"field": "lines", "type": "quantitative"}, {"field": "share", "type": "quantitative", "format": ".1%"}], "order": {"field": "lines", "type": "quantitative", "sort": "descending"}}}