{ "schema": "AI_VILLAGE_S3_ORION_FINAL_ANALYSIS_V1", "published_date": "2026-09-23", "session_id": "S3-20-20260923T195817+0200-717cc5", "public_url": "https://www.t-1-t.com/ai/village/roundtables/s3-20-models/", "epistemic_status": "POST_SESSION_ANALYSIS_NO_VOTE_NO_WINNER", "session": { "duration_minutes": 60, "portances": 16, "candidates": 20, "attempts": 133, "direct_responses": 79, "original_errors": 54, "genuine_late_recovered": 15, "authentic_textual_completions": 94, "semantic_replays_separate": 5 }, "hardware": { "S3": { "hostname": "t-node-s3", "cpu": "Intel Core i7-8565U @ 1.80GHz", "threads": 8, "ram_gib": 7.6, "swap_gib": 7.9, "gpu": "Intel UHD Graphics 620", "disk_free_gib": 432, "os": "Debian GNU/Linux 13", "ollama_installed": false, "source": "live T-LINUX scan 2026-09-23" }, "Orion3": { "hostname": "PREDATUT", "cpu": "Intel Core i7-12700F", "logical_processors": 20, "ram_gib": 31.8, "free_ram_gib_at_check": 17.1, "gpu": "NVIDIA GeForce RTX 3070", "gpu_vram_mib": 8192, "gpu_vram_free_mib_at_check": 4794, "driver": "572.16", "source": "live T-LINUX/Orion scan 2026-09-23" } }, "findings": [ "S3 doit être traité comme un nœud CPU sobre et spécialisé, pas comme un petit Orion3.", "Le couple RAG + architectures efficaces/atypiques est la piste la plus structurante pour S3.", "Orion3 est le terrain naturel des expériences multimodales, raisonnement et modèles plus lourds.", "La diversité linguistique/culturelle doit être une capacité recherchée explicitement, pas un simple sous-produit des benchmarks.", "La séance a montré une contagion factuelle autour d’un faux S3/A100; les prochaines séances doivent verrouiller un bloc de faits matériels non négociable.", "Les mentions explicites de candidats sont des indices d’attention seulement. Plusieurs portances ont dérivé hors format; aucune absence de mention ne vaut rejet.", "Les 5 replays postérieurs sont conservés séparément et ne sont pas traités comme des réponses originales de la séance." ], "action_baskets": { "S3_INFRA_RAG": [ "C09", "C10" ], "S3_GENERATIVE_BENCHMARK": [ "C01", "C02", "C03", "C04", "C05", "C06" ], "S3_RD_DIVERSITY": [ "C07", "C08" ], "BRIDGE_S3_ORION": [ "C11" ], "ORION_MULTIMODAL_REASONING": [ "C12", "C13", "C14", "C15" ], "ORION_RD_DIVERSITY_HEAVY": [ "C16", "C17", "C18", "C19", "C20" ] }, "candidate_attention_note": "Counts are lexical mentions in 94 authentic textual completions; they are not support scores.", "candidates": [ { "id": "C01", "name": "LiquidAI LFM2.5-1.2B-Instruct", "kind": "generative", "size": "1.2B", "architecture": "hybrid short-convolution + attention", "proposed_target": "S3", "why": "Architecture très différente des Transformers classiques; edge/on-device; longue fenêtre; multilingue.", "discussion": "Licence LFM 1.0 plutôt qu'Apache/MIT.", "official_url": "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct", "post_session_status": "BENCHMARK_S3", "role": "Edge/génératif", "post_session_rationale": "Très compact, architecture LFM distincte et positionnement edge. Bon test de portance légère sur CPU.", "key_uncertainty": "Mesurer latence CPU, RAM réelle, français et qualité conversationnelle; licence LFM 1.0 à conserver visible.", "explicit_mentions_authentic": 8, "distinct_portances_authentic": 6, "phases_with_mentions": { "INDEPENDENT_REVIEW": 5, "TECHNICAL_FIT": 3 } }, { "id": "C02", "name": "Microsoft BitNet b1.58-2B-4T", "kind": "generative", "size": "2B", "architecture": "native 1.58-bit BitNet", "proposed_target": "S3", "why": "Très basse précision native; intérêt mémoire/énergie; architecture/runtime atypiques.", "discussion": "Principalement anglais; bitnet.cpp spécifique.", "official_url": "https://huggingface.co/microsoft/bitnet-b1.58-2B-4T", "post_session_status": "BENCHMARK_S3_RD", "role": "Efficacité / architecture", "post_session_rationale": "Cas scientifique majeur pour S3 : 1.58-bit natif et runtime BitNet spécifique.", "key_uncertainty": "Intégration bitnet.cpp distincte du chemin Ollama; mesurer débit, mémoire et stabilité sur i7-8565U.", "explicit_mentions_authentic": 3, "distinct_portances_authentic": 2, "phases_with_mentions": { "TECHNICAL_FIT": 3 } }, { "id": "C03", "name": "Google RecurrentGemma-2B-IT", "kind": "generative", "size": "2B", "architecture": "recurrent", "proposed_target": "S3", "why": "Architecture récurrente; comparaison intéressante avec LFM/RWKV.", "discussion": "Conditions Gemma; modèle ancien mais architecturalement distinct.", "official_url": "https://huggingface.co/google/recurrentgemma-2b-it", "post_session_status": "BENCHMARK_S3_SECONDARY", "role": "Récurrence", "post_session_rationale": "Contrepoint récurrent compact utile face à LFM/BitNet/RWKV.", "key_uncertainty": "Modèle plus ancien et conditions Gemma; intérêt surtout comparatif.", "explicit_mentions_authentic": 0, "distinct_portances_authentic": 0, "phases_with_mentions": {} }, { "id": "C04", "name": "IBM Granite 4.0 H-Micro", "kind": "generative", "size": "3B", "architecture": "hybrid Mamba-2 / Transformer", "proposed_target": "S3 / Orion3", "why": "Hybride compact; candidat scientifique pour local/low latency.", "discussion": "Runtime Mamba2 et quantification à tester sur 8 Go RAM.", "official_url": "https://www.ibm.com/granite/docs/models/granite4-0", "post_session_status": "BENCHMARK_S3_RD", "role": "Hybride Mamba-2", "post_session_rationale": "3B hybride conçu pour usages locaux / basse latence; bonne expérience d’architecture.", "key_uncertainty": "Support Mamba-2 et quantification à vérifier sur S3; ne pas inférer la vitesse avant benchmark.", "explicit_mentions_authentic": 0, "distinct_portances_authentic": 0, "phases_with_mentions": {} }, { "id": "C05", "name": "Hugging Face SmolLM3-3B", "kind": "generative", "size": "3B", "architecture": "small-scale reasoning/chat Transformer", "proposed_target": "S3 / Orion3", "why": "Petite échelle ouverte; multilingue; famille absente de T-LINUX.", "discussion": "Moins radical architecturalement que LFM/BitNet/RWKV.", "official_url": "https://huggingface.co/HuggingFaceTB/SmolLM3-3B", "post_session_status": "BENCHMARK_S3_SECONDARY", "role": "Compact polyvalent", "post_session_rationale": "Petit modèle ouvert et multilingue, utile comme baseline moderne simple.", "key_uncertainty": "Moins différenciant architecturalement que LFM, BitNet ou RWKV.", "explicit_mentions_authentic": 0, "distinct_portances_authentic": 0, "phases_with_mentions": {} }, { "id": "C06", "name": "LG AI EXAONE 4.0-1.2B", "kind": "generative", "size": "1.2B", "architecture": "reasoning/non-reasoning Transformer", "proposed_target": "S3", "why": "Famille coréenne; on-device; raisonnement et outils.", "discussion": "EN/KO/ES; licence EXAONE spécifique, observée NC/recherche-éducation pour cette release.", "official_url": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B", "post_session_status": "BENCHMARK_S3_WITH_LICENSE_GUARD", "role": "Diversité coréenne / on-device", "post_session_rationale": "1.2B, nouvelle famille culturelle et technique, pertinente pour la diversité du Village.", "key_uncertainty": "Licence EXAONE spécifique/NC observée pour cette release; périmètre d’usage à respecter.", "explicit_mentions_authentic": 0, "distinct_portances_authentic": 0, "phases_with_mentions": {} }, { "id": "C07", "name": "Lelapa AI InkubaLM-0.4B", "kind": "generative", "size": "0.4B", "architecture": "small Llama-like", "proposed_target": "S3", "why": "Très forte valeur de diversité africaine et faible ressource.", "discussion": "Capacité limitée; CC-BY-NC-4.0.", "official_url": "https://huggingface.co/lelapa/InkubaLM-0.4B", "post_session_status": "RD_S3_DIVERSITY", "role": "Diversité africaine", "post_session_rationale": "Très petit modèle apportant une diversité linguistique/culturelle absente du parc.", "key_uncertainty": "Capacité limitée et licence CC-BY-NC; intérêt d’abord expérimental, pas remplacement d’un assistant généraliste.", "explicit_mentions_authentic": 1, "distinct_portances_authentic": 1, "phases_with_mentions": { "DIVERGENCE_AND_MINORITY": 1 } }, { "id": "C08", "name": "RWKV7 Goose 2.9B — 20260831", "kind": "generative", "size": "2.9B", "architecture": "attention-free recurrent RWKV-7", "proposed_target": "S3 / Orion3", "why": "État récurrent constant; coût/token constant; architecture très différente.", "discussion": "Checkpoint de base, pas assistant aligné final; plutôt R&D/post-training.", "official_url": "https://huggingface.co/RWKV/RWKV7-G1j-2.9B-20260831", "post_session_status": "RD_PRIORITY_S3_ORION", "role": "Architecture récurrente RWKV", "post_session_rationale": "Candidat le plus explicitement discuté; architecture attention-free et état récurrent constant très différente.", "key_uncertainty": "Checkpoint de base : post-training/alignment nécessaire avant d’en faire une portance conversationnelle.", "explicit_mentions_authentic": 12, "distinct_portances_authentic": 4, "phases_with_mentions": { "TECHNICAL_FIT": 5, "DIVERGENCE_AND_MINORITY": 1, "DIVERSITY_AND_ARCHITECTURE": 4, "REVISED_POSITIONS": 1, "OPEN_QUESTIONS": 1 } }, { "id": "C09", "name": "BAAI BGE-M3", "kind": "embedding", "size": "~568M class", "architecture": "multilingual dense+sparse+ColBERT retrieval", "proposed_target": "S3 — RAG", "why": "Cœur RAG autonome riche; dense, sparse et multi-vector dans un modèle.", "discussion": "Plus lourd qu'Arctic pour simple embedding.", "official_url": "https://huggingface.co/BAAI/bge-m3", "post_session_status": "INFRASTRUCTURE_S3_RAG", "role": "RAG riche", "post_session_rationale": "Peut faire de S3 un nœud RAG autonome : dense + sparse + multi-vector.", "key_uncertainty": "Plus lourd qu’un embedding simple; mesurer coût CPU/indexation et bénéfice réel des modes multiples.", "explicit_mentions_authentic": 0, "distinct_portances_authentic": 0, "phases_with_mentions": {} }, { "id": "C10", "name": "Snowflake Arctic-Embed-M-v2.0", "kind": "embedding", "size": "305M class", "architecture": "multilingual embedding", "proposed_target": "S3 — RAG", "why": "RAG léger; 74 langues; contrepoint efficace à BGE-M3.", "discussion": "Moins polyvalent que BGE-M3 multi-mode.", "official_url": "https://huggingface.co/Snowflake/snowflake-arctic-embed-m-v2.0", "post_session_status": "INFRASTRUCTURE_S3_RAG", "role": "RAG léger multilingue", "post_session_rationale": "Alternative plus légère à BGE-M3, 74 langues, Apache-2.0.", "key_uncertainty": "Comparer qualité/latence/mémoire à BGE-M3 sur le corpus réel d’AI^VILLAGE.", "explicit_mentions_authentic": 3, "distinct_portances_authentic": 2, "phases_with_mentions": { "INDEPENDENT_REVIEW": 1, "DIVERGENCE_AND_MINORITY": 1, "DIVERSITY_AND_ARCHITECTURE": 1 } }, { "id": "C11", "name": "Mistral Ministral 3 3B Instruct 2512", "kind": "generative_multimodal", "size": "3.4B LM + 0.4B vision encoder", "architecture": "Mistral 3 edge multimodal", "proposed_target": "S3 / Orion3", "why": "Vision, français, agentique, edge; plus récent que Mistral 7B actuel.", "discussion": "Famille Mistral déjà présente; quantification nécessaire sur S3.", "official_url": "https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512-BF16", "post_session_status": "BENCHMARK_ORION_THEN_S3", "role": "Multimodal edge", "post_session_rationale": "Vision + texte, français, famille edge; pont naturel entre S3 et Orion3.", "key_uncertainty": "Commencer sur Orion3; tenter S3 uniquement après quantification et mesure réelle de RAM.", "explicit_mentions_authentic": 0, "distinct_portances_authentic": 0, "phases_with_mentions": {} }, { "id": "C12", "name": "Google Gemma 3n E4B-it", "kind": "generative_multimodal", "size": "8B raw / ~4B footprint class", "architecture": "MatFormer selective activation/offload", "proposed_target": "Orion3", "why": "Texte+image+vidéo+audio; >140 langues; architecture imbriquée.", "discussion": "Intégration complexe; conditions Gemma; recoupe services multimédias Orion.", "official_url": "https://huggingface.co/google/gemma-3n-E4B-it", "post_session_status": "BENCHMARK_ORION", "role": "Multimodal large diversité", "post_session_rationale": "Texte, image, vidéo et audio avec forte couverture linguistique; apporte perception au Village.", "key_uncertainty": "Intégration plus complexe et conditions Gemma; éviter de dupliquer des services multimédias sans cas d’usage.", "explicit_mentions_authentic": 2, "distinct_portances_authentic": 1, "phases_with_mentions": { "DIVERGENCE_AND_MINORITY": 1, "DIVERSITY_AND_ARCHITECTURE": 1 } }, { "id": "C13", "name": "Microsoft Phi-4-mini-flash-reasoning", "kind": "generative_reasoning", "size": "compact", "architecture": "SambaY/GMU hybrid reasoning", "proposed_target": "Orion3", "why": "Raisonnement maths/code spécialisé; efficacité et latence.", "discussion": "Principalement anglais; documenté surtout pour raisonnement mathématique.", "official_url": "https://huggingface.co/microsoft/Phi-4-mini-flash-reasoning", "post_session_status": "BENCHMARK_ORION_REASONING", "role": "Raisonnement spécialisé", "post_session_rationale": "Architecture Phi Flash atypique et focalisation maths/code; utile comme spécialiste.", "key_uncertainty": "Ne pas extrapoler ses résultats maths à un assistant généraliste; runtime custom à valider.", "explicit_mentions_authentic": 6, "distinct_portances_authentic": 3, "phases_with_mentions": { "INDEPENDENT_REVIEW": 2, "TECHNICAL_FIT": 3, "DIVERGENCE_AND_MINORITY": 1 } }, { "id": "C14", "name": "Microsoft Phi-4-multimodal-instruct", "kind": "generative_multimodal", "size": "5.6B class", "architecture": "Phi multimodal text+vision+audio", "proposed_target": "Orion3", "why": "Portance perceptive texte+image+audio; 24 langues.", "discussion": "Compréhension multimodale, pas générateur image/audio.", "official_url": "https://huggingface.co/microsoft/Phi-4-multimodal-instruct", "post_session_status": "BENCHMARK_ORION_MULTIMODAL", "role": "Perception texte/vision/audio", "post_session_rationale": "Portance perceptive potentielle, complémentaire des modèles texte actuels.", "key_uncertainty": "Compréhension multimodale ≠ génération image/audio; vérifier VRAM et pipeline.", "explicit_mentions_authentic": 3, "distinct_portances_authentic": 1, "phases_with_mentions": { "TECHNICAL_FIT": 1, "DIVERSITY_AND_ARCHITECTURE": 2 } }, { "id": "C15", "name": "Ai2 OLMo 3 7B Think", "kind": "generative_reasoning", "size": "7B", "architecture": "OLMo 3 reasoning", "proposed_target": "Orion3", "why": "Recherche ouverte avec code/checkpoints/détails; maths/code.", "discussion": "Anglais; 7B proche de modèles Orion existants.", "official_url": "https://huggingface.co/allenai/Olmo-3-7B-Think", "post_session_status": "RD_ORION", "role": "Recherche ouverte / raisonnement", "post_session_rationale": "OLMo apporte une famille de recherche très ouverte et un spécialiste de raisonnement.", "key_uncertainty": "7B recoupe plusieurs modèles Orion existants; intérêt à démontrer par benchmark différentiel.", "explicit_mentions_authentic": 3, "distinct_portances_authentic": 3, "phases_with_mentions": { "INDEPENDENT_REVIEW": 3 } }, { "id": "C16", "name": "NVIDIA Nemotron Nano 9B v2", "kind": "generative", "size": "9B", "architecture": "Nemotron-H hybrid", "proposed_target": "Orion3", "why": "Nouvelle famille NVIDIA; profil conversation/outils; hybride.", "discussion": "Quantification/offload nécessaires sur RTX 3070 8GB; licence NVIDIA.", "official_url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2", "post_session_status": "RD_ORION", "role": "Hybride NVIDIA", "post_session_rationale": "Nouvelle famille et architecture Nemotron-H, potentiellement intéressante pour outils/conversation.", "key_uncertainty": "9B + custom_code : quantification/offload à mesurer sur RTX 3070 8GB.", "explicit_mentions_authentic": 5, "distinct_portances_authentic": 4, "phases_with_mentions": { "INDEPENDENT_RECOVERY": 1, "TECHNICAL_FIT": 4 } }, { "id": "C17", "name": "OpenBMB MiniCPM4-8B", "kind": "generative", "size": "8B", "architecture": "MiniCPM4", "proposed_target": "Orion3", "why": "Famille OpenBMB absente; contrepoint chinois/anglais à Qwen.", "discussion": "custom_code et runtime local à valider.", "official_url": "https://huggingface.co/openbmb/MiniCPM4-8B", "post_session_status": "RD_ORION", "role": "Diversité OpenBMB", "post_session_rationale": "Nouvelle famille chinoise/anglaise qui diversifie au-delà de Qwen.", "key_uncertainty": "Runtime custom_code et bénéfice par rapport aux modèles déjà présents à démontrer.", "explicit_mentions_authentic": 1, "distinct_portances_authentic": 1, "phases_with_mentions": { "INDEPENDENT_RECOVERY": 1 } }, { "id": "C18", "name": "inclusionAI Ling-3.0-tiny", "kind": "generative_reasoning_moe", "size": "7.9B total / 1.3B active", "architecture": "3:1 KDA–MLA hybrid + sparse MoE", "proposed_target": "Orion3", "why": "Architecture 2026 très originale; raisonnement hybride et faible coût actif.", "discussion": "Runtime récent; intégration à valider.", "official_url": "https://huggingface.co/inclusionAI/Ling-3.0-tiny", "post_session_status": "RD_PRIORITY_ORION", "role": "MoE hybride très atypique", "post_session_rationale": "7.9B total / 1.3B actifs, architecture KDA–MLA + MoE; forte valeur de diversité architecturale.", "key_uncertainty": "Runtime récent/custom; tester INT4 et stabilité avant toute intégration permanente.", "explicit_mentions_authentic": 1, "distinct_portances_authentic": 1, "phases_with_mentions": { "DIVERGENCE_AND_MINORITY": 1 } }, { "id": "C19", "name": "Cohere Labs Aya Expanse 8B", "kind": "generative_multilingual", "size": "8B", "architecture": "Command-family multilingual research model", "proposed_target": "Orion3", "why": "Forte diversité linguistique/culturelle.", "discussion": "CC-BY-NC + Acceptable Use Policy; périmètre recherche.", "official_url": "https://huggingface.co/CohereLabs/aya-expanse-8b", "post_session_status": "RD_ORION_DIVERSITY", "role": "Multilingue / culturel", "post_session_rationale": "Aya apporte une diversité linguistique forte et une philosophie de recherche multilingue distincte.", "key_uncertainty": "Licence CC-BY-NC + AUP; poids 8B. L’API Cohere a été retirée en 2026 mais les poids de recherche restent publiés séparément.", "explicit_mentions_authentic": 1, "distinct_portances_authentic": 1, "phases_with_mentions": { "REVISED_POSITIONS": 1 } }, { "id": "C20", "name": "OpenAI gpt-oss-20b", "kind": "generative_reasoning_moe", "size": "21B total / 3.6B active", "architecture": "open-weight sparse/MoE", "proposed_target": "Orion3 — expérimental lourd", "why": "Gros modèle local, raisonnement configurable, agentique/outils, Apache 2.0.", "discussion": "8GB VRAM insuffisants pour tout GPU; offload/CPU à mesurer.", "official_url": "https://developers.openai.com/api/docs/models/gpt-oss-20b", "post_session_status": "RD_HEAVY_ORION", "role": "Raisonnement MoE / agentique", "post_session_rationale": "21B total / 3.6B actifs : bon test de limite supérieure locale, outils et raisonnement.", "key_uncertainty": "RTX 3070 8GB insuffisante pour un chargement GPU complet; offload/CPU et latence doivent être mesurés.", "explicit_mentions_authentic": 4, "distinct_portances_authentic": 1, "phases_with_mentions": { "INDEPENDENT_REVIEW": 1, "REVISED_POSITIONS": 1, "OPEN_QUESTIONS": 1, "DIVERGENCE_AND_MINORITY": 1 } } ], "next_experiment": "Run small isolated benchmarks by capability basket before any permanent installation; record RAM/VRAM, latency, energy, runtime complexity, language behavior and failure modes." }