Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
Use the Instruct models for general instruction following in French and English, after testing them on your own task. The card says the model should be tested before it enters an industrial pipeline. It recommends further training for specific uses and a RAG pipeline for knowledge (ev-card-instruct).
Code generation and maths need extra training, according to the card. The base model is a next-word predictor and must be fine-tuned before end use (ev-card-base). No document sets out forbidden uses, and no acceptable-use policy ships.
Known limitations, bias & failure modes
All items come from the publisher's cards and dataset cards (ev-card-instruct, ev-posttrain-data, ev-pretrain-data).
- The model hallucinates and can memorise little at its size.
- Post-training used almost entirely English data, although the base saw about 30% French.
- Post-training sequences were 16,384 tokens, against a 131k base window.
- Tool calling is marked work in progress.
- Web and public-domain sources carry toxic and historically biased text, and filtering is described as ongoing.
- Much of the post-training data was generated by Qwen3 and other third-party open-weights models, which the dataset card says can introduce linguistic and cultural bias.
Openness tier & components
Luciole 1.1 Instruct is at the open_weights_recipe tier, with a ceiling of 4 on the openness dimension.
| Component | Grade | Evidence |
|---|---|---|
| Weights | open | safetensors, ungated, Apache-2.0 (ev-hf-platform) |
| Training data | partial | pretraining corpus and post-training data are public with per-source licences; three English web subsets sit on an external server that was not tested (ev-pretrain-data, ev-posttrain-data) |
| Training code | open | full pipeline in one GPL-3.0 repository, including NeMo pretraining and NeMo-RL fine-tuning recipes (ev-code-repo) |
| Evaluation | partial | lighteval-based scripts and task lists are public; results are published as figures (ev-code-repo, ev-bench) |
| Documentation | open | detailed cards and dataset cards; no technical report (ev-report-absent) |
| Licence | open | Apache-2.0 declared (ev-card-instruct, ev-notice) |
The base card also lists intermediate checkpoints and training logs. The logs are in the Base repo. The checkpoints sit on a Linagora file server whose listing could not be opened (ev-card-base).
Licence terms & permitted use
The Instruct weights are Apache-2.0, ungated, with commercial use allowed. The card says the model is released "under the Apache License 2.0" with copyright held by LINAGORA SAS. The NOTICE file in the code repository names LINAGORA SAS as owner of the models (ev-card-instruct, ev-notice).
The Instruct repos contain no LICENSE file. The Base repos contain a LICENSE.md that holds the Apache-2.0 text with the copyright line left blank (ev-licence-file). Apache-2.0 section 4 expects a copy of the licence to travel with redistributed copies, so add the text yourself.
The training code is GPL-3.0 (ev-code-licence). Copying or distributing the code requires GPL-3.0 for the derivative. The weights are separately licensed and the NOTICE file does not extend the GPL to them.
The post-training data was generated by Qwen3, Ministral-3 and, through NVIDIA source sets, GPT-OSS-120B and Qwen3-235B. Their Hugging Face tags show Apache-2.0 and no output restriction was found. The licence files of Ministral-3 were not opened (ev-upstream).
The dataset licences differ from the weights. Both datasets are CC-BY-SA-4.0 and say each source keeps its own licence.
Supply-chain & provenance
The canonical source is huggingface.co/OpenLLM-France, which Hugging Face lists as not verified, with 20 members. The 23B Instruct repo holds ten safetensors shards, no pickle file and no custom code. Hugging Face flagged no file, and it lists LFS hashes. The publisher gives no separate checksum file and no signature. That is about 4 of 8 checklist boxes (ev-hf-platform).
The Instruct models descend from Luciole-23B-Base, which the cards link as the base model. The base card gives an initialiser range, a Nemotron3-22B architecture recipe from NVIDIA's NeMo framework and no parent checkpoint. It does not say the model was trained from scratch. The data points to random initialisation, but the cards do not state it (ev-card-base).
Independence from other models is limited. The pretraining mix includes Qwen3-8B reformulations of FineWeb 2 and Wikipedia, and model-generated NVIDIA and OpenThoughts sets (ev-pretrain-data). The post-training data is largely Qwen3 output, and the chat template is modelled on Qwen3's (ev-posttrain-data, ev-card-instruct).
The 8B uses a hybrid Mamba-transformer and its repos ship Python modelling code. Review that code before running it (ev-siblings).
EU AI Act posture
Luciole is a general-purpose AI model. The base card gives about 4.9 trillion tokens and 576,587 H100 GPU hours, which is about 6.8e23 FLOPs by the 6 x parameters x tokens estimate. That is below the 1e25 systemic-risk threshold, so Article 55 does not apply (ev-card-base).
Apache-2.0 would normally support the Article 53 open-source exemption, and weights, architecture and usage information are public. The exemption never covers a copyright policy or a public training-content summary.
The dataset cards list each source with its licence, describe robots.txt filtering and PII removal, and offer an opt-out form. This covers most of a training-content summary, in a format other than the AI Office template. No separate copyright policy, Annex XI document or technical report exists (ev-pretrain-data, ev-report-absent). Code of Practice signatory status is not stated. A fine-tuner who places a derivative on the EU market may become a provider for it. This is guidance, not legal advice.
Benchmarks & evaluation
Every figure is the publisher's. The Instruct card shows a bar chart against four 22B to 24B instruct models (ev-bench).
| Measure | Luciole-23B-Instruct-1.1 | Reference points in the figure |
|---|---|---|
| Overall averaged score | about 64 | Mistral-Small-3.2 about 67.5, Mistral-Small-2501 about 62, EuroLLM-22B about 59, Gaperon-24B about 41 |
| ifeval / ifeval-fr | about 0.83 / 0.74 | first of five on both |
| GSM8K | about 0.82 | Mistral-Small-3.2 about 0.92 |
| ifbench test / multiturn | about 0.32 / 0.42 | Mistral-Small-3.2 about 0.35 / 0.49 |
| mixeval easy / hard | about 0.74 / 0.53 | Mistral-Small-3.2 about 0.84 / 0.64 |
Values are read off the bars and are approximate. The comparison set is the publisher's choice. The Thinking-1.1 card claims comparability with Olmo-3-32B-Think-DPO on English tasks and the best French results among the thinking models tested. The base card gives French, English, multilingual, translation and RULER figures against Gaperon-24B, EuroLLM-22B and Mistral-Small-24B-Base. No independent re-run was found.
Score summary
Rubric 1.1, computed by scripts/score.py: openness 4, provenance 2, legal 3, safety 1, performance 2, operational 3, governance 3. The reasons are in each dimension rationale in entry.yaml.
Ownership verdict
Level: partial. Use and modify, transparency and data control are strong. Reliability is weak because safety is rated 1. See the ownership block in entry.yaml.
Contradictions and surprises
- The Thinking card documents
enable_thinking=False, so its thinking can be switched off. - The cards and the consortium site say the data is open-licensed, yet the pretraining dataset card lists Claire sources under CC BY-NC-SA 4.0 and one source with no stated licence under a single CC-BY-SA-4.0 label. The cards do not say whether Claire entered the training mix.
- The site says datasets are republished for auditability. Three English web subsets are on a separate server, and the Base card says checkpoints use revision tags, but Hugging Face lists only the main branch.
- The base card does not say "from scratch", and both data mixes contain output of other models.
- The training code is GPL-3.0 and the weights are Apache-2.0.
- The Modelfile comment in the GGUF repo names "Luciole-1B-Instruct-1.2", which matches no published repo.
Follow-up entries
Four releases differ from this entry on a scored attribute and need their own entries.
- Luciole Base (23B, 8B, 1B): has a LICENSE.md, no post-training and no safety tuning, and the base cards list eight languages.
- Luciole-23B-Thinking-1.1: a safety-alignment note, a different post-training branch and 346 downloads.
- Luciole-8B-KD-Math (Base and Thinking-1.1): pruned and distilled from Luciole-23B-Base.
- Luciole Audio-1.0 (1B, 8B, 23B): built on Instruct-1.1 and Parakeet, trained on a CC-BY-NC-SA audio dataset under an Apache label.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
StrongApache-2.0 is declared for the Instruct weights, which are ungated and fine-tunable, with commercial use allowed. The Instruct repos have no LICENSE file, and LINAGORA SAS is named as owner in the NOTICE file (ev-card-instruct, ev-notice).
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
StrongWeights, training and evaluation code, the pretraining corpus and the post-training data are public, and the cards are detailed. Gaps are the unpublished technical report and the external server for three web subsets (ev-pretrain-data, ev-posttrain-data, ev-code-repo).
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.