Skip to main content

​
A European model, strengthened and optimized by LightOn’s R&D team

Alfred Sv5 Pn Alfred-sv5 is not an off-the-shelf model. It is built on a modern 24-billion-parameter architecture, extensively fine-tuned by LightOn’s R&D team, with a full retraining of the LoRA components that support LightOn’s RAG pipeline.
The result is improved stability, more reliable outputs on enterprise data, and seamless integration within search, retrieval, and enterprise agent workflows. This new version also delivers a major fix to an issue observed in Alfred 4.2, which could occasionally trigger instabilities or looping behaviors. Alfred-Sovereign 5 now provides significantly enhanced stability, especially for long, complex, or highly structured queries.

​
Technical elements without making them the core topic

In a mature RAG architecture, the key to quality lies in the retrieval pipeline and orchestration, not in the model spec sheet. Still, for teams who need technical touchpoints, Alfred-sv5 offers:
  • a context window of up to 32,000 tokens,
  • an architecture refined to improve instruction-following and reduce repetitive or looping outputs,
  • strong support for function calling and structured outputs, enabling tighter integration with tools, agents, and enterprise systems.
These capabilities complement rather than replace the upstream work: indexing, embeddings, retrieval, filtering, and context composition.

​
Sovereignty vs performance: a clear choice, soon to be expanded

Alfred-Sovereign 5 addresses a very real market demand: some organizations specifically require a European-origin LLM, even when other open-source models already provide complete confidentiality. To serve these different expectations transparently, LightOn is structuring its model offering along two product lines:
  • Alfred-Sovereign: for teams that prioritize European origin, and alignment with internal political or regulatory frameworks.
  • Alfred-Frontier (coming soon): for organizations that prioritize raw performance and the highest possible quality, regardless of the model’s origin.

​
A component of a RAG system not the destination

In a fully operational RAG system, the LLM is just one component among many: indexing, embeddings, retrieval, guardrails, agent workflows, and post-processing all play decisive roles.
Alfred-Sovereign 5 positions itself as the sovereign core model within this broader pipeline without pretending to be the sole driver of system performance. This combination a well-designed RAG pipeline + a model fine-tuned for enterprise use + sovereign or high-performance options is what enables organizations to balance political requirements, regulatory constraints, and real-world impact.