As AI consultants based in Rosenheim, Germany, we use this page to put Aleph Alpha in context: the Heidelberg company that started out with its own language models in 2019, shifted its focus to the PhariaAI enterprise platform in 2024, has been preparing a merger with Cohere since September 2026, and on 3 October 2026 released Kolibri-1, once again a model of its own with an open licence. This page is based on the model cards on Hugging Face, the PhariaAI documentation, the company website and the Cohere merger announcements. As of 4 October 2026.
Aleph Alpha at a glance
| Headquarters | Heidelberg; further locations per the website: Berlin, Bayreuth, Munich |
| Founded | 2019 by Jonas Andrulis and Samuel Weinbach |
| Leadership | Ilhan Scheer (CEO), Samuel Weinbach and Yasser Jadidi (Co-Chief Research Officers) |
| Investor | Schwarz Group and Dieter Schwarz Foundation as co-leads of the November 2023 financing round |
| References per the website | Federal Employment Agency, State of Baden-Württemberg, Federal Ministry for Digital and Transport, Deutsche Bank, SAP, Infineon, Hella |
| Current model | Kolibri-1 (3 October 2026), Apache 2.0 |
| Platform | PhariaAI v1.260900.0 (Assistant, Studio, OS, Catch, Government Assistant) |
| Corporate status | Definitive merger agreement with Cohere of 16 September 2026, closing pending |
As a German company, Aleph Alpha is subject to German and European law. Per its website, the company positions itself for “mission critical environments” in government, automotive, financial services and semiconductor manufacturing; its core promise is operation on infrastructure the customer controls.
Model portfolio: from Luminous to Kolibri-1
| Model | Year | Architecture | Context | Licence | Status |
|---|---|---|---|---|---|
| Kolibri-1 | Oct 2026 | MoE, 78.1B / 3.46B active | 1,048,576 | Apache 2.0 | current, recommended |
| Pharia-1 LLM-7B | 2024 | dense, 7.04B | 8,192 | Open Aleph License (non-commercial) | current, PhariaAI default model |
| T-Free HAT (Llama 3.1 8B/70B) | 2025 | tokenizer-free on Llama backbone | as Llama 3.1 | Open Aleph License | research |
| Luminous Base/Extended/Supreme | 2022 to 2024 | dense, up to ~70B | 2,048 | proprietary | no longer marketed |
Luminous was the first model family, offered through a proprietary API from German data centres and with explainability features. In September 2024, Aleph Alpha shifted its focus from the model business to the PhariaAI platform; Luminous has not been marketed to new customers since, and the former developer documentation redirects to the PhariaAI documentation. Pharia-1 LLM-7B appeared the same year as a dense 7-billion-parameter model for seven European languages. Important for context: while the weights are published on Hugging Face, the Open Aleph License permits only non-commercial research and educational use. Production use runs through an agreement with Aleph Alpha, typically via PhariaAI.
With the T-Free models (January 2025, 70B variant September 2025), Aleph Alpha Research published a tokenizer-free architecture: a character-level encoder and decoder wrap a word-level backbone based on Llama 3.1. Aleph Alpha reports better compression and adaptability to new languages as well as advantages on German benchmarks; the model card describes the release as an early research result with non-optimised inference. Kolibri-1 does not continue this architecture but uses a conventional UniBPE tokenizer with 128,000 entries.
Kolibri-1 has been the current model since 3 October 2026: a German-English Mixture-of-Experts transformer trained from scratch with 78.1 billion parameters (3.46 billion active), 1 million tokens of context, four reasoning levels, tool calling and an Apache 2.0 licence. Per the tech report, its overall scores in German (70.8) and English (75.5) are ahead of the open MoE models in its class, with weaknesses in closed-book knowledge and multi-turn tool calling. Architecture, benchmarks, hardware and operations are covered in detail on the Kolibri-1 page.
PhariaAI: the enterprise platform
PhariaAI has been Aleph Alpha’s main product since 2024 and, per the documentation, is currently available in version v1.260900.0. Its building blocks:
- PhariaAssistant: chat interface with ready-to-use tools and workflows for business users.
- PhariaStudio: development environment in which teams build, test and evaluate AI applications.
- PhariaOS: operations layer for deploying and scaling workloads on your own infrastructure.
- PhariaCatch: capture and preservation of knowledge from human-machine collaboration.
- Pharia Government Assistant (version PGA 2025.08.20): variant for public authorities.
The documentation names Pharia-1 LLM-7B as the model. The platform is designed for operation in your own data centre or in a sovereign cloud; a partnership with STACKIT has existed since the 2023 financing round and, per Cohere’s announcement, is to form the “technical backbone” of the joint offering. Whether and when Kolibri-1 will be documented as a model in PhariaAI was not apparent as of 4 October 2026.
Merger with Cohere (April to September 2026)
On 24 April 2026, Canadian AI company Cohere and Aleph Alpha announced their planned combination; on 16 September 2026, per their joint announcement, the definitive merger agreement followed. The key points per the announcements:
- Brand and structure: the combined company operates globally as Cohere, dual-headquartered in Berlin and Toronto; Heidelberg remains a research centre. More than 1,000 employees across both continents.
- Leadership: on closing, Aleph Alpha CEO Ilhan Scheer becomes Cohere’s Chief Operating Officer and co-founder Samuel Weinbach Chief Research Officer; Aidan Gomez remains CEO.
- Valuation and investor: media reports consistently cite a combined valuation of about USD 20 billion; per Cohere, the Schwarz Group is committing USD 600 million to Cohere’s upcoming Series E round. Per the April announcement, Aleph Alpha shareholders hold about 10 percent of the combined company.
- Positioning: the “first transatlantic sovereign AI solution” for governments and regulated industries in Europe and Canada, delivered on STACKIT.
- Product roadmap: in April, a unified Command-Pharia 1 model was announced for Q4 2026; the September announcement points to further product details to come. Kolibri-1 was released independently under the Aleph Alpha Research name.
- Closing: expected later in 2026, subject to regulatory approvals in both jurisdictions.
For customers, until closing: both companies operate as separate legal entities and existing contracts remain in place. For new projects we recommend actively following the model and platform roadmap of the combined company, in particular the relationship between Kolibri-1 and Cohere’s Command models.
Data protection, sovereignty and operations
Aleph Alpha’s sovereignty claim rests on three points: the company is based in Germany, the models are trained on European infrastructure (Kolibri-1 in Germany and Finland per the tech report), and operation takes place on the customer’s infrastructure. For Kolibri-1, the tech report also describes a data pipeline that, with the EU AI Act, the GPAI Code of Practice and the GDPR in mind, filters illegal, harmful and pirated content and redacts personal data.
Three points to check before deciding:
- Licence: only Kolibri-1 is under Apache 2.0. Pharia-1 and the T-Free models may be used non-commercially only without an agreement with Aleph Alpha.
- Deployment path: Aleph Alpha does not offer new customers a public pay-per-token API. The choice is between self-hosting the open weights, PhariaAI in your own data centre and the announced STACKIT offering.
- Hardware: Kolibri-1 needs about 78 GB of GPU memory in FP8 (two H100s or one H200/B200), Pharia-1 runs on a single GPU.
| Model | GPU memory | Example configuration |
|---|---|---|
| Kolibri-1 (FP8) | ~78 GB | 2× H100 SXM5 / 2× A100 80 GB / 1× H200 / 1× B200 |
| Pharia-1 LLM-7B | ~14 GB | 1× L40S / RTX 4090 / A10 |
Use cases
Aleph Alpha addresses organisations that process sensitive data under regulation:
- Public administration: federal, state and municipal authorities (references per the website: Federal Employment Agency, State of Baden-Württemberg, Federal Ministry for Digital and Transport).
- Industry: automotive suppliers, semiconductors, drive technology; the Kolibri tech report cites its own industry RAG evaluations for these sectors.
- Financial sector: banks and insurers with residency and audit obligations.
- Aerospace: one of Kolibri-1’s target sectors per the tech report.
Integration with CompanyGPT
Kolibri-1 can be connected to CompanyGPT as an OpenAI-compatible vLLM endpoint and run behind the AI Gateway: employees get a familiar chat interface while prompts and documents stay on your own infrastructure. Pharia-1 is suitable for this only with a commercial agreement.
Our recommendation
For new projects, Kolibri-1 is the relevant option from the Aleph Alpha portfolio: Apache 2.0, German language quality within its class, 1 million tokens of context and a deployment path that gets by with two H100s or one B200. Before deploying, check the quality in multi-turn tool calling and on factual questions without retrieval; on both points the model trails comparison models per the vendor’s own figures. Pharia-1 LLM-7B remains sensible for single-GPU environments or existing PhariaAI installations, but commercially requires an agreement with Aleph Alpha. Luminous should only be kept running in existing projects.
Anyone who needs image input, broad language coverage or a hosted API with EU data residency should compare Mistral as a European alternative or OpenAI GPT via Microsoft Foundry’s EU data zones. Given the ongoing merger with Cohere, we recommend structuring contracts with the combined company’s product roadmap in mind. For an assessment of which option fits your requirements, contact innFactory AI Consulting.