Pharia-1-LLM-7B
Multilingual 7B parameter decoder model by Aleph Alpha Research
Aleph Alpha Research released the Pharia-1-LLM-7B family, comprising two seven-billion parameter models. These autoregressive transformers support multiple European languages and are optimised for industrial use cases like engineering. The weights are available for research under a specific open licence.
- Contains 7 billion parameters in a decoder-only architecture
- Optimised for German, French, and Spanish languages
- Available under Open Aleph License for research use
- Designed for automotive and engineering domain tasks
- Offers both aligned and unaligned model variants
What it is
The Pharia-1-LLM-7B family consists of two foundation models created by Aleph Alpha Research. Each variant contains seven billion parameters and operates as an autoregressive decoder. The system is designed to generate concise responses suitable for domain-specific tasks in sectors such as automotive and engineering. It supports several European languages including German, French, and Spanish. The models aim to provide efficient performance for applications requiring length-controlled outputs.
How it came about
Aleph Alpha Research developed these models to complement existing open-source options. The team focused on creating a system that balances speed with cultural and linguistic optimisation for key European markets. The release includes both a base control model and an aligned version. The company states that training data was curated to comply with national laws and copyright regulations within the European Union.
Inside the model
These models use a decoder-only transformer architecture with rotary position embeddings. They are trained on next token prediction using a multilingual corpus. The training data covers English, German, French, Spanish, Italian, Portuguese, and Dutch. One variant, the control model, remains unaligned to offer flexibility for developers who wish to add their own safety layers. The second variant undergoes additional alignment training to reduce risks in chat-based interactions. Both versions prioritise token efficiency.
What it can do
The models perform text generation, classification, summarisation, and question answering. They are tested primarily in English, German, Spanish, and French. The control variant is engineered to match the performance of other leading models in the seven to eight billion parameter range. It delivers concise answers which helps in critical applications where predictable output length is valuable. The aligned version is recommended for direct user interaction due to its safety adjustments.
How to use it
Weights are hosted on Hugging Face under the Open Aleph License. This permit allows non-commercial research and educational use. Commercial users can access the full stack via on-premise installation or cloud deployment through direct contact with the developer. An Intelligence Layer SDK is also available for approved accounts to help build evaluation pipelines. The company does not store prompt data or train on user inputs from its public API.
Safety and testing
The aligned model includes specific training to mitigate usage risks. The unaligned control model requires developers to implement their own guardrails for safe deployment. Prohibited uses include illegal activities, terrorism, and generating malicious code. The company monitors for policy violations and may terminate accounts or take legal action if terms are breached. An appeals mechanism exists for reported violations.
Questions people ask about Pharia-1-LLM-7B
What languages does Pharia-1-LLM-7B support?
The model was trained on English, German, French, Spanish, Italian, Portuguese, and Dutch. It is specifically optimised for German, French, and Spanish. Testing was conducted primarily in English, German, Spanish, and French to ensure quality in these key markets.
Can I use this model for commercial products?
The weights on Hugging Face are limited to non-commercial research and education. For commercial use, you must contact Aleph Alpha Research directly. They offer on-premise installation and full stack access for customers who need to deploy the model in production environments.
What is the difference between the control and aligned versions?
The control model is unaligned and offers flexibility for developers to add custom safety measures. The aligned version has undergone extra training to reduce risks and is recommended for chat applications. Both have seven billion parameters and similar core capabilities.
Good reading and sources
There is no English Wikipedia article on Pharia-1-LLM-7B yet, so the facts on this page come from the sources listed above: huggingface.co/Aleph-Alpha/Pharia-1-LLM-7B-control. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.