LINAGORA AI
Digital commons

Luciole: open language models, in three sizes

Luciole is not a product we rent to you, it is a digital common. The weights are published under the Apache 2.0 licence, at 1, 8 and 23 billion parameters, to cover needs that have nothing in common with one another. The models are developed by LINAGORA and the OpenLLM France consortium, with the support of BPI France under France 2030.

https://huggingface.co/OpenLLM-France

Why Luciole matters

Weights nobody can take away from you
Apache 2.0 on all three sizes. You download, you audit, you retrain, you redistribute. No permission to ask for
A European digital common
Developed by LINAGORA and the OpenLLM France consortium, with the support of BPI France under France 2030. The model belongs to the community that maintains it
An open chain end to end
Base models and instruction-tuned models, quantised builds, documented training method. That is what makes AI Act compliance demonstrable rather than declarative
A French-language foundation
The base models were trained on a substantial share of French data, around thirty per cent, which is rare in an open model of this size
A documented limitation
The 1.1 instruction-tuned versions were post-trained almost entirely on English data. We say so because it is written on the model cards, and because a common is judged on what it documents
Built into our other blocks
Luciole works natively with Open-RAG for corpus querying and with LinTO for speech, and it can be served as a sovereign managed service if you would rather not operate it yourself

Three sizes, three uses

The right size is not the largest, it is the smallest that meets your requirements. All three models share the same licence and the same context window.

Luciole-1B-Instruct-1.1
Parameters
1 billion
Context window
16,384 tokens
Licence
Apache 2.0

What it is for

The lightest. Quantised GGUF builds let it run on modest hardware, with Ollama for instance. Its size limits what it can memorise: keep it for bounded tasks, backed by a document base.

Luciole-8B-Instruct-1.1
Parameters
8 billion
Context window
16,384 tokens
Licence
Apache 2.0

What it is for

The balance. Instruction following across varied tasks: maths, science, code, general chat, corpus querying and translation. This is the default size for an internal assistant served on reasonable infrastructure.

Luciole-23B-Instruct-1.1
Parameters
23 billion
Context window
16,384 tokens
Licence
Apache 2.0

What it is for

The most capable. Post-trained in three phases, one of them with thinking traces, then aligned through direct preference optimisation. For long reasoning and difficult synthesis.

The base models, also published, accept context windows of up to 131,000 tokens. The instruction-tuned versions were trained on sequences of 16,384 tokens.

We can operate this building block for you

You can deploy it yourself, it is open. You can also hand it to us: we run it to your requirements, on platforms subject neither to the Cloud Act nor to any other extraterritorial law.

Discover managed inference

Run Luciole yourself, or hand it to us