Aleph Alpha

💰 Pricing & Plans

🔧 API & Developer Experience

• •OpenAI Responses API compatibility: POST {base_url}/v1/responses with an AA_TOKEN bearer header. Point the OpenAI Python SDK, PydanticAI, LangGraph (langchain-openai with use_responses_api=True) or plain curl at your deployment URL and the same code runs unchanged. The Stateful Responses layer adds conversation persistence via previous_response_id, instruction inheritance across turns and SSE streaming.
• •Two kinds of tool calling: Function tools run client-side: the model returns a function_call, your code executes it and posts the result back as function_call_output. MCP tools run server-side: Stateful Responses connects to an MCP endpoint, executes the tool inside the agentic loop and feeds the result back, so the inference backend never sees the tool call itself.
• •Agentic retrieval with inline citations: A file_search tool searches uploaded documents or configured knowledge-base stores through the Sherlock document service; a web_search tool queries live pages through LinkUp. Both are agentic — the model decides whether and when to search — and both return url_citation annotations mapping [1] and [2] markers back to the exact source passages.
• •Guardrails in the request path: Every request is screened before it reaches the LLM. Three providers ship by default: llama-guard-3 (LLM classifier, roughly 500ms-2s per check), regex (deterministic PII and blocklist patterns, microseconds) and llm-judge (a natural-language policy the application manager writes). Policies are defined per checkpoint — input or output only — and the pipeline short-circuits on the first violation.
• •Background jobs for long reasoning: POST /v1/responses with background:true returns 202 Accepted and a req_ job id immediately. Poll GET /v1/responses/{req_id} until the status leaves in_progress, then use the permanent resp_ id for follow-up operations. Built for multi-step agentic runs where holding a synchronous connection open is impractical.
• •Skills, sharing and authorisation: Versioned instruction bundles follow the open Agent Skills specification — a SKILL.md manifest plus optional scripts, references and assets — uploaded to /v1/skills and attached to a request through a hosted shell tool. Sharing across users, organisations or tenants is enforced with OpenFGA, a Zanzibar-style relationship engine, rather than in application code.
• •Official SDKs and a self-host serving path: The PhariaInference SDK, PhariaStudio SDK and PhariaData SDK are on PyPI (pharia-inference-sdk 0.1.6, pharia-studio-sdk 0.1.7, pharia-data-sdk 0.1.6; Python 3.10-3.12). PhariaEngine compiles your own code into WebAssembly skills via componentize-py. For local Kolibri serving, aleph-alpha-inference 1.0.0 is the official vLLM plugin.

🇩🇪 Sovereign AI Stack: Kolibri + PhariaAI

Aleph Alpha ships a full sovereign stack rather than a single endpoint. Kolibri is a mixture-of-experts model with 78B total parameters but only 3.5B active per token, trained end-to-end in Germany on a bilingual corpus that is roughly 62.5% English and 23.9% German, with a tokenizer tuned to German word structure. Its signature feature is controllable reasoning effort - none, low, medium or high - letting one deployment trade latency for accuracy per request; at low effort the released model keeps 98% of its high-effort score while using only 76% of the completion tokens. Around it, PhariaOS runs agents from WASM-compiled skills behind approval gates with every action logged, and PhariaStudio adds evaluation datasets and traces. Because Kolibri ships under Apache 2.0, the whole stack can run on your own GPUs.

🌐 China Access & Latency

Aleph Alpha serves no mainland-China region and publishes no ICP filing, so for Chinese teams the answer is not slower but generally unavailable. Aleph Alpha's own model documentation states that customers receive identical models regardless of geographic location and input language, but that availability remains subject to sanction regimes, technology export regulations and other applicable restrictions — an EU-sovereign vendor whose flagship deployments are European public administration and defence. Measured from an Alibaba Cloud Malaysia host, aleph-alpha.com answers with roughly 740ms TTFB (Bayreuth, Deutsche Telekom AS3320) and api.aleph-alpha.com returns 401 in about 800ms with a 260ms TCP connect; the documentation site is far quicker at roughly 46ms because it sits on Vercel edge nodes. The practical route is to keep everything inside China: PhariaAI installs from Helm charts onto your own Kubernetes, and Kolibri's Apache 2.0 weights can be served locally with the aleph-alpha-inference vLLM plugin, which removes the cross-border data path entirely.