{"resourceId":"linagora-luciole-nemotron-openllm-france-20260708","versions":[{"version":"external-ceb851cef0a1f6d0ff6a2b469047f8f69bdc8ae1adc2a60bfac98102bc26912b","resource":{"id":"linagora-luciole-nemotron-openllm-france-20260708","title":"LINAGORA: Luciole open models adapt the Nemotron recipe, trained on Jean Zay","organization":"LINAGORA / OpenLLM France","sector":"Software vendor and open-model consortium","geography":"France","publishedAt":"2026-07-08","publicationDate":"2026-07-08","eventDate":"2026-07-08","sourceName":"Carnet souverain (Michel-Marie Maudet blog, reposting LINAGORA)","sourceLabel":"Vendor-affiliated blog","sourceUrl":"https://blog.maudet.cloud/blog/linagora-nvidia-nemotron-luciole/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["developers-agents","governance-procurement"],"finding":"The post says LINAGORA and OpenLLM France build open multilingual models with GENCI and AI Factory France support on the Jean Zay supercomputer. Luciole comes in 1B, 8B (hybrid Mamba-Transformer) and 23B versions, all Apache 2.0, with checkpoints, training data and code published and a corpus of around 5 trillion tokens. All Luciole architectures are adapted from Nemotron, and model cards credit a custom adaptation of the Nemotron 3 Nano recipe and training with NeMo on Jean Zay H100s. Twake.ai is listed in the same open-source ecosystem.","sledRelevance":"Moderate for state AI offices and universities considering open models built on public compute; explains the Nemotron link behind NVIDIA's Twake.ai claim.","evidence":"Vendor-affiliated description; no benchmarks.","architectureImplications":"Nemotron-derived architectures and NeMo tooling; deployable where data stays under the operator's control.","governanceImplications":"Openness beyond weights supports auditability.","securityPrivacyImplications":"Not addressed.","caveats":"Why now: it is the documented basis for NVIDIA's Oct 13 'built on Nemotron' attribution; it does not say which models Twake.ai deployments run.","streamIds":["nvidia"],"roles":{"sales":"Sales framing: LINAGORA's July 8 post documents the actual NVIDIA link: its Luciole open models (1B, 8B, 23B) adapt the Nemotron 3 Nano training recipe and were trained with NeMo on France's public Jean Zay supercomputer. Twake.ai is named as part of the same open-source ecosystem. Interpretation: NVIDIA's role is in model development tooling, a narrower claim than 'built on Nemotron'.","engineering":"Engineering framing: all Luciole models use architectures adapted from Nemotron, are Apache 2.0, and ship checkpoints, training data and code, with a roughly 5 trillion token multilingual corpus. Pre-sales engineers can cite this as an example of public compute (GENCI, IDRIS) used to build auditable open models, while confirming which model a given Twake.ai deployment actually runs.","delivery":"Delivery framing: the post describes foundation models, not finished assistants, so public deployments need fine-tuning and evaluation work. Delivery teams should plan model evaluation for each use case. Acceptance criteria below are proposals."},"roleDetails":{"sales":{"customerProblem":"Public-sector buyers need to know what 'built on' a vendor's model actually means for provenance and control.","stakeholders":["State AI governance offices","University AI research groups","Public-sector procurement officers"],"discoveryQuestions":["Do you require model provenance documentation?","Would open training data and checkpoints satisfy your audit needs?","Do you have public compute for fine-tuning?"],"valueHypothesis":"Interpretation: fully open models with documented lineage can ease audits; the post reports no task benchmarks.","potentialEngagement":"Bounded model-provenance review for a public AI deployment, distinguishing training recipe, datasets and runtime model.","unsupportedClaims":["That Twake.ai deployments run Luciole or Nemotron","Benchmark performance","Production readiness of foundation models"]},"engineering":{"fit":"Moderate for research universities and state AI offices interested in open models trained on public compute.","architectureIntegration":"Luciole 1B/8B/23B; hybrid Mamba-Transformer 8B; trained with NeMo on Jean Zay H100s.","prerequisites":["Fine-tuning compute","Evaluation datasets"],"constraints":["Foundation models, not chat assistants","European-language focus"],"security":"Open data and code aid auditing; no security evaluation reported.","proofOfValue":["Evaluate a Luciole model on a representative public-sector task","Trace dataset licenses"]},"delivery":{"work":"Fine-tune and evaluate open models for a specific public use case.","dependencies":["Public compute allocation","Evaluation team"],"ownership":"Research or AI platform team.","skillsAdoption":"Requires ML engineering skills for fine-tuning and evaluation.","governanceCheckpoints":["Model provenance documented","Evaluation completed before deployment"],"acceptanceCriteria":["Proposed: task accuracy measured against a baseline model"],"risks":["Overstating vendor involvement","Insufficient evaluation"]}},"retrievedAt":"2026-10-14T02:56:44Z","enrichedAt":"2026-10-14T03:01:25Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Public compute access lets regional builders participate.","procurementImplications":"Apache 2.0 models reduce license constraints.","operatingModelImplications":"Regional model builders depend on public supercomputing allocations.","sourceVerification":{"openedUrl":"https://blog.maudet.cloud/blog/linagora-nvidia-nemotron-luciole/","referenceExcerpt":"All Luciole models have architectures adapted from Nemotron.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}