Ethically trained AI startup Pleias launches new small models for reflection optimized for rags with built -in quotes

French AI startup Playle made waves at the end of last year with Starting your ethically trained pleias 1.0 family from small language patterns – Among the first and only so far, which are fully built on the scraping of “open” data, that is, the data explicitly marked as public, open code or unlicensed and not protected with copyright.

The company now has announced the publication Of the two small -scale reflection models with open source, designed specifically to generate extraction (RAG), citation synthesis and a structured multilingual output.

The launch includes two main models-Pleias-RAG-350M and Pleias-RAG-1B-Each also available in the GGUF-optimized format on CPU, which makes a total of four options ready for implementation.

All of them are based on Pleias 1.0 and can be used independently or in conjunction with other LLMs that the organization can now either plan to unfold. It seems that all are available under an Apache 2.0 open -open license, which means they are Eligible for organizations to take, change and deploy commercial use.

Rag, as you will remember, is the widely used technique that businesses and organizations can unleash to hook AI large language (LLM) as Openai’s GPT-4O., Google’s Flash Gemini 2.5., Anthropic’s Claude Sonnet 3.7 or Coher’s commandOr open source alternatives such as Llama 4 and Deepseek V3 to external knowledge bases, such as corporate documents and cloud warehouses.

This is often necessary for businesses that want to build chatbots and other AI applications that refer to their internal policies or product catalogs (an alternative that drives a long LLM context with all the necessary information, may not be suitable for use of businesses that are concerned about security and costs).

The family of the Pleias-RAG model is the latest effort to overcome the gap between accuracy and efficiency in small language models.

These models are aimed at businesses, developers and researchers seeking profitable alternatives to large -scale language models without compromising tracking, multilingual opportunities or structured work motives.

The target user base is actually the home continent of Europe of Pleias, as co -founder Alexander Doria told Venturebeat via a direct message on the social network X:

“The main motivation is the difficulty of scaling RAG applications in Europe. Most private organizations have small graphic processors (may have changed, but not long ago less than 2% of all (NVIDIA) H100 (GPU) were in Europe). However, there is a strong stimulus for self-hosting reasons, including GDP.

“SLMS have progressed significantly in the last year, but they are still too often conceived as “mini-jet” and we see a significant decline in non-English efficiency, both in terms of source understanding and in terms of text generation. So we are pleased to achieve most of our goals:

• A real alternative to 7-8b RAG models even a processor and other limited infra.
• Fully verifiable models coming with citation support.
• Preserving European language results. “

However, of course, models that are open code under the Apache 2.0 license means that anyone can take and use them freely anywhere in the world.

Focused on grounding, quotes and facts

The main feature of the new Pleias-Rag models is their native support for citing sources with literal quotes, fully integrated in the process of the model’s conclusions.

Unlike the methods of quoting after hock or external pipelines, Pleias-RAG models generate quotes directly using a syntax inspired by the Wikipedia reference format.

This approach allows shorter, more quotation fragments while maintaining a check.

Cital grounding plays a functional role in adjustable settings.

For sectors such as healthcare, legal and finance-when decision-making must be documented and traceable and traceable inquiries offer a direct way to audit. Pleias positions this choice of design as an ethical imperative, which is aligning in accordance with the growing regulatory requirements for explanatory AI.

Proto agent?

Pleias-RAG models are described as “protoangenic” they can autonomously assess whether the application is understandable, determine whether it is trivial or complex, and decide whether to respond, reformulate or refuse based on the adequacy of the source.

Their structured outcome includes reports of language detection, applications and sources analysis and a reasoned response.

Despite its relatively small size (Pleias-RAG-350M has only 350 million parameters), models have a behavior traditionally associated with larger, agent systems.

According to Pleias, these opportunities stem from a specialized medium training pipeline, which combines synthetic generation of data with iterative prompts for reasoning.

Pleias-RAG-350M is explicitly designed for a limited environment. It performs well on standard processors, including mobile infrastructure.

According to the internal indicators, the unpunished version of GGUF produces full outputs for reasoning for approximately 20 seconds at 8GB RAM settings. Its small imprint puts it in a niche with very few competitors, such as QWEN-0.5 and Smollm, but with a much stronger focus on the structured source synthesis.

Competitive performance within tasks and languages

In comparison estimates, Pleias-RAG-350M and Pleias-RAG-1B outperform the most open models under 4 billion parameters, including Llama-3.1-8B and QWen-2.5-7b, on tasks such as Hotpotqa, 2wikimultihopqa and Musique.

These repeated comparative rag indicators test the model’s ability to think in multiple documents and to identify distracting-general requirements in enterprise knowledge systems.

The power of the models extends to multilingual scenarios. With translated indicators to compare French, German, Spanish and Italian, Pleias models show a slight deterioration of productivity.

This sets them apart from other SLMs, which usually experience a loss of performance of 10-35% when working with non-English requests.

Multilingual maintenance stems from a careful design of a tokeneizer and synthetic competition, which includes exercises to switch the language. Models not only discover the language of a user request, but seek to respond in the same language – an important feature for global implementation.

In addition, Doria emphasized how models can be used to increase the performance of other existing models that an entity can already use:

“We plan to use the models in orchestration setup, especially since their computing costs are low. Many interesting results from the assessment: even the 350m model turned out to be good in completely different answers from the answers (Meta) Llama and (alibaba) QWEN, so there is a real complaint, we are making a real complaint. refer to to attribute qwen.… ”

Open access and licensing

According to Doria and Technical book Details of the training of the Pleias-Rag family, the models were trained on: “A common body to create a training set of rags (all 3 million examples came from it). We used (Google) Gemma on top to generate synthetic traces of reasoning, as the license allows re -use/retraining.”

Both models are placed under the Apache 2.0 license, which allows the use of commercial reuse and integration into larger systems.

Pleias emphasizes the suitability of models for integration into assistants trained in search, educational instruments and user support systems. The company also provides an API library for simplifying the structured formatting of entrance-exit for developers.

The release of models is part of a broader boost of Pleias to reposition small LLMs as tools for structured reasoning, not as conversational common-purpose bots.

With the use of external memory architecture and systematic citation methods, the Pleias-RAG series offers a transparent, audit alternative to the less transparent border models.

Prospect

Looking forward, Pleias plans to expand the capabilities of the models by longer processing of the context, the stronger integration of the search and the adjustment of the individual for a more consistent representation of identity.

Reinforcement of training, especially in domains as a citation accuracy, is also investigated, where the test verification can be measured algorithmic.

The team also actively collaborates with partners such as the Wikimedia Foundation to support targeted search integration using reliable sources.

Ultimately, the current use of RAG-specific conversions, models and work processes may be eliminated, as more modern AI models are trained and deployed, those that include a one-off use of rags and agents. As Doria told Venturebeat via DM:

“In the long run, my conviction is that both the classic RAG gas pipelines and the long context models will be violated by the search agents. We started moving in this direction: Therefore, the model is already equipped with many features that are currently externalized in RAG applications (request reformulation, rerance, etc.). Obviously, we strive to continue and integrate the search capacity and the capacity to process sources directly into the model itself. My conviction is that RAG will disappear in some way as it is automated by agent models that can direct their own work processes.“

With Pleias-RAG-350M and 1B, the company relies that small models are paired with strong skeletal reflections and verifiable results-they can compete with much larger colleagues, especially in multilingual and infrastructure restrictions.