Aleph Alpha vs. DeepSeek: data protection per plan

not publicly documented (no public DPA text; the privacy notices cover only the website, no product/customer contract; when retrieved on 2026-10-18 the documentation link "Data privacy" leads to the website privacy notices, and the support portal is accessible only after logging in)

not publicly documented

not publicly documented

not publicly documented: The privacy notices apply to the website aleph-alpha.com. Regarding customer data in PhariaAI operation, there is no statement on training on the platform, documentation or press release pages. The platform page states: "You retain full control over your data and processes." (translated) (no statement on training). Kolibri training data: according to the press release of 2026-10-05, checked against a blocklist of more than 4.5 million URLs; this concerns the origin of the model training data, not customer data. Technical Report on Kolibri (q13): "we redact PII from all training sources" (likewise concerns model training data, not customer data).

Terms of use (as of 2026-03-27, section 4.3): DeepSeek may use inputs and outputs 'to a minimal extent', after encryption and de-identification, to improve the services and the underlying technologies; opt-out available by switching off 'Improve the model for everyone'. Privacy policy: data is used, among other things, to train and improve the models; right to object to use for training. Document 'Model Mechanism and Training Methods of DeepSeek': in the optimisation phase, a small portion of the training data may be based on user input, which is then encrypted, de-identified and anonymised.

The Open Platform terms (effective 2026-04-29) govern ownership of inputs and outputs (the customer retains rights to inputs and receives rights to outputs) but contain no express statement on the use of API inputs for training; they refer to the document 'Model Principles and Training Methods Explanation', which likewise does not distinguish between chat and API. The terms of use (section 4.3) and the privacy policy apply to 'Services' in general; an API-specific opt-out is not described. For end-user data in its own applications, according to the privacy policy and the Open Platform terms (5.5), the developer is the controller.

Platform page: "Available on-premise, air-gapped or in a sovereign SaaS solution" (translated); models are provided "in sovereign cloud, hybrid and on-premises environments" (translated). Press release 2025-05-26: PhariaAI-as-a-Service on STACKIT (Schwarz Group), "hosted entirely on European infrastructure". PhariaAI launch 2024-08-26: "available as SAAS solutions or on-premise in the customers' own datacenter"; hybrid execution (on-prem + cloud) possible. Specific data centre locations per operating model, and whether any operating mode runs exclusively in Germany: not publicly documented. Note: when retrieved on 2026-10-14, the model page (q9) contains the statement "Ownership, development and supply chain lie entirely in Germany and the EU" (translated) and "Confidential data always remains with the customer" (translated) (no information on data centre locations); the previously recorded wording "No foreign jurisdiction has access" (translated) is no longer contained there. Release notes PhariaAI v1.260900.0 (2026-09-03, q14): "Availability: On-premise and hosted". Technical Report (q13) on model training (not on customer operation): "train it on infrastructure in Germany and Finland".

Privacy policy (as of 2026-02-10): 'we directly collect, process and store your Personal Data in People's Republic of China'. The category 'US only/EU option' does not apply; the selection list does not include China, hence 'not publicly documented' as the enum value here. An EU or US region: not publicly documented.

Privacy policy (as of 2026-02-10): 'we directly collect, process and store your Personal Data in People's Republic of China'. The category 'US only/EU option' does not apply; the selection list does not include China, hence 'not publicly documented' as the enum value here. An EU or US region: not publicly documented. Region of API processing: not publicly documented.

• not publicly documented (no list published). Named as infrastructure partner for PhariaAI-as-a-Service: STACKIT (Schwarz Group), press release 2025-05-26.

not publicly documented (no public plan; contact via "Contact sales" (translated))

free of charge (website app description: 'Chat with leading AI models for free.'; no paid chat plan found)

Per 1M tokens, in USD (pricing page): deepseek-flash: input (cache hit) USD 0.003 off-peak / USD 0.006 peak; input (cache miss) USD 0.15 / USD 0.3; output USD 0.6 / USD 1.2. deepseek-v4-pro: cache hit USD 0.022 / USD 0.044; cache miss USD 0.66 / USD 1.32; output USD 1.98 / USD 3.96. Peak hours: 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday (excluding Chinese public holidays); otherwise half price. Search and Fetch API USD 0.75 each per 1,000 successful requests. Billed against topped-up or granted credit; DeepSeek may change prices; no indication whether incl. VAT.

yes (PhariaAI stack according to the documentation: PhariaAssistant, PhariaStudio, PhariaOS, PhariaCatch; models including Kolibri, available since 2026-10-03)

yes (AI chatbot; according to the terms of use, services based on large language models)

yes (models deepseek-flash = DeepSeek-V4.1-Flash-1017 and deepseek-v4-pro = DeepSeek-V4-Pro-0813, each with 1M-token context, up to 384K output, thinking/non-thinking mode, tool calls, JSON output, Responses API, OpenAI- and Anthropic-compatible interface; vision only with deepseek-flash; additionally Search and Fetch API)