Hub Dollars · couverture ≥30B

18 octobre 2026 · Coût de calcul équivalent en location · R&D humaine et datasets exclus

Le total du Hub reste inconnu. Les montants par organisation couvrent une sélection de familles. Le registre est une exploration de scénarios par dépôt : ses lignes ne doivent pas être additionnées. La couverture ≥30B dépend des métadonnées et n’est pas exhaustive des modèles réellement publiés.

Le calcul partagé, au fil des publications

Cumul des 75 postes sélectionnés, selon les mois de sortie : 991 M$ au terme du scénario central. Ce n’est ni le total exhaustif du Hub, ni une série des dépenses historiques. Les tarifs et hypothèses V4 restent fixes.

Chronologie provisoire. Certaines dates sont vérifiées dans les annonces de publication des poids ; les autres sont des dates de publication Epoch ou des mois de travail encore à vérifier. Une annonce d’API ou de papier peut précéder les poids. Le tableau ci-dessous expose cette distinction.

Organisation Scénario de coût

Mois

Courbes lissées entre les valeurs mensuelles pour la lecture ; les points intermédiaires ne sont pas des observations.

Trait bleu : chaque budget de famille est compté au dernier mois retenu. Pointillé vert : au premier mois. La zone entre les deux montre la sensibilité au calendrier d’allocation ; elle ne représente pas une fourchette de coûts ou un intervalle de confiance. Pas de ventilation inventée entre versions. Les budgets et suppléments non ventilés héritent de la même fenêtre ; les mois provisoires peuvent encore déplacer les deux traits.

Graphique SVG · PNG · Série mensuelle CSV · Dates et sources CSV · JSON détaillé

Voir les 75 fenêtres de publication et leur provenance

Organisation / famille	Fenêtre mensuelle	Central (M$)	Provenance des dates
Meta Llama 2	2023-07 → 2023-07	7.26	Llama 2-70B : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Meta Llama 3 / 3.1 / 3.2 / 3.3	2024-04 → 2024-12	110.87	Llama 3-70B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Llama 3.3 70B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Meta Llama 4 Scout + Maverick	2025-04 → 2025-04	20.30	Llama 4 Scout : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Llama 4 Maverick : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Alibaba Qwen 2.5	2024-09 → 2025-02	30.44	Qwen2.5-72B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Qwen2.5-VL-72B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Alibaba Qwen 3	2025-04 → 2025-09	41.96	Qwen3-235B-A22B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Qwen3-Next-80B-A3B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Alibaba Qwen 3.5	2026-02 → 2026-02	28.30	Qwen3.5 397B-A17B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Qwen3.5-122B-A10B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Alibaba Qwen 3.8	2026-08 → 2026-08	38.48	Qwen3.8-2.4T-A95B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
DeepSeek DeepSeek V2	2024-05 → 2024-09	2.11	DeepSeek-V2 (MoE-236B) : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. DeepSeek-V2.5 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
DeepSeek DeepSeek V3 / R1 / V3.x	2024-12 → 2025-12	8.36	DeepSeek-V3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. DeepSeek-V3.2 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
DeepSeek DeepSeek V4 Pro + Flash	2026-04 → 2026-04	26.35	DeepSeek-V4-Pro : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. DeepSeek-V4-Flash : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
DeepSeek DeepSeek V4.1 Flash	2026-09 → 2026-09	9.72	DeepSeek V4.1 Flash : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Moonshot Kimi K2 / K2.5	2025-07 → 2026-02	12.65	Kimi K2 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Kimi K2.5 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Moonshot Kimi K3	2026-07 → 2026-07	42.12	Kimi K3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Meituan LongCat Flash + 2.0	2025-09 → 2026-06	29.82	LongCat-Flash : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. LongCat-2.0 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Ant Group Ling 1T / 2.6	2025-10 → 2026-04	18.38	Ling-1T : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Ling-2.6-flash : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Z.ai GLM 4.5 / Air	2025-07 → 2025-12	13.17	GLM-4.5 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. GLM-4.7 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Z.ai GLM 5 / 5.x	2026-02 → 2026-08	15.39	GLM-5 : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. GLM-5.3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Tencent Hunyuan Large / Hy3 / Hy4	2024-11 → 2026-08	23.41	Hunyuan-Large : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Tencent Hy4 preview : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
LG AI Research K-EXAONE 1 / 2	2025-12 → 2026-07	10.95	K-EXAONE : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. K-EXAONE 2.0 : Publication Epoch (proxy) — source 1 · source 2 · source 3 · source 4 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
NVIDIA Nemotron 3 / 3.5	2025-12 → 2026-10	21.25	Nemotron 3-Nano-30B-A3B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Nemotron 3.5 Lightning : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Nemotron 3.5 Super-VL allowance : Non résolu — borne au mois du relevé — source 1 · source 2 · source 3 · source 4 · source 5 Release month not independently resolved. Late edge extended to the snapshot month; not an asserted October release.
MiniMax MiniMax Text-01 / M1	2025-01 → 2025-06	10.84	MiniMax-Text-01 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. MiniMax-M1-80k : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
MiniMax MiniMax M2 / M3	2025-10 → 2026-06	11.14	MiniMax-M2 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. MiniMax-M3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Baidu ERNIE 4.5 text / VL	2025-06 → 2025-06	7.56	@2025-06-29 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
StepFun Step 5	2026-09 → 2026-09	9.11	Step 5 Preview : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Mistral AI Mixtral 8x7B / 8x22B	2023-12 → 2024-04	6.45	Mixtral 8x7B : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Mixtral 8x22B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Mistral AI Mistral Medium 3.5	2024-11 → 2026-04	24.88	@2024-11-01 : Mois de travail — à vérifier — source 1 · source 2 · source 3 · source 4 · source 5 · source 6 Curated working month for the budgeted release; open-weight availability still requires independent verification. Mistral Medium 3.5 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
TII Falcon 7B / 40B / 180B	2023-05 → 2023-09	14.01	@2023-05-01 : Annonce primaire — source 1 First open release May 2023; earlier Epoch announcement is not the weight release. Falcon-180B : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
BigScience BLOOM	2022-07 → 2022-07	2.26	BLOOM-176B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Black Forest Labs FLUX 1 / 2	2024-08 → 2025-11	8.00	Flux.1 [dev] : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. FLUX.2 : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Stability AI Stable Diffusion 1 / 2 / XL / 3 / 3.5	2022-08 → 2024-10	5.00	@2022-08-22 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2024-10-29 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Alibaba Wan 2.1 / 2.2	2025-02 → 2025-08	3.00	Wan 2.1 14B I2V : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. @2025-08-26 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
OpenAI Whisper	2022-09 → 2023-11	0.83	Whisper : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. @2023-11-06 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Google Gemma 2	2024-06 → 2024-07	5.53	@2024-06-27 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2024-07-31 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Google Gemma 3	2025-03 → 2025-03	7.00	@2025-03-12 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Google Gemma 4	2026-04 → 2026-06	9.14	Gemma 4 31B IT : Publication Epoch (proxy) — source 1 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Gemma 4 12B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
OpenAI gpt-oss 120B / 20B	2025-08 → 2025-08	5.78	@2025-08-05 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Ai2 OLMo 2	2024-12 → 2025-03	3.54	OLMo 2 Furious 7B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. OLMo 2 32B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
IBM Granite 3 dense	2024-10 → 2024-12	1.49	@2024-10-21 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2024-12-18 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Microsoft Phi 4	2025-01 → 2025-01	1.78	@2025-01-08 : Annonce primaire — source 1 January 8 update announces Hugging Face availability; December commits do not establish public access.
NVIDIA Nemotron 4 / H	2024-06 → 2025-04	53.41	Nemotron-4 340B : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. @2025-04-14 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Arcee / Prime Intellect Trinity Large	2026-02 → 2026-02	2.74	Arcee Trinity Large : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
StepFun Step Video	2025-02 → 2025-02	8.49	@2025-02-24 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Thinking Machines Inkling / Small	2026-07 → 2026-07	32.20	Inkling : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Inkling-Small : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Xiaomi MiMo 7B / V2 Flash / V2.5 Pro	2025-06 → 2026-04	22.21	@2025-06-05 : Mois de travail — à vérifier — source 1 · source 2 · source 3 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2026-04-23 : Mois de travail — à vérifier — source 1 · source 2 · source 3 Curated working month for the budgeted release; open-weight availability still requires independent verification.
China Telecom TeleChat 1 / 2	2024-04 → 2024-09	15.84	@2024-04-01 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2024-09-20 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Swiss AI Apertus 8B / 70B	2025-09 → 2025-09	15.44	@2025-09-02 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Upstage Solar Open 100B / Open2 250B	2025-12 → 2026-06	12.55	@2025-12-31 : Mois de travail — à vérifier — source 1 · source 2 · source 3 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2026-06-28 : Mois de travail — à vérifier — source 1 · source 2 · source 3 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Databricks DBRX	2024-03 → 2024-03	5.37	@2024-03-27 : Mois de travail — à vérifier — source 1 · source 2 Curated working month for the budgeted release; open-weight availability still requires independent verification.
xAI Grok 1	2024-03 → 2024-03	6.00	@2024-03-17 : Annonce primaire — source 1 Open release March 17, 2024, not November 2023 API debut.
Cohere Command R / R+	2024-03 → 2024-04	9.01	@2024-03-11 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2024-04-04 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Mistral AI Mistral Large 3 — base	2025-12 → 2025-12	7.97	Mistral Large 3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
NVIDIA Nemotron Nano 2 — 12B base	2025-08 → 2025-08	2.66	NVIDIA-Nemotron-Nano-12B-v2 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
NVIDIA Cosmos 1 / Predict1 — allocation scenario	2025-01 → 2025-01	27.00	Cosmos-Predict1-7b-Video2World : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability. Cosmos-Predict1-14b-Video2World : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
NVIDIA Cosmos 3 Nano + Super — generator increment	2026-05 → 2026-05	9.72	@2026-05-31 : Annonce primaire — source 1 NVIDIA lists May 31, 2026 for Nano and Super.
Cohere Command A — 111B base	2025-03 → 2025-03	21.58	Cohere Command A : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Cohere Command A+ — 25B active base	2026-05 → 2026-05	6.75	Command A+ : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
xAI Grok 2 — 2024 base scenario	2025-08 → 2025-08	9.76	@2025-08-23 : Historique des uploads (proxy) — source 1 August 2025 weights/upload history; 2024 is the training/API generation, not open release. Commit time alone does not prove public visibility.
SK Telecom A.X K1	2025-12 → 2025-12	6.42	A.X K1 : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
SK Telecom A.X K2 — pretraining	2026-07 → 2026-07	2.92	A.X K2 : Publication Epoch (proxy) — source 1 · source 2 · source 3 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Alibaba Qwen 2 — dense 0.5/1.5/7/72B	2024-06 → 2024-06	6.15	@2024-06-07 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Alibaba Qwen 2 57B-A14B — upcycling increment	2024-06 → 2024-06	0.68	@2024-06-07 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Z.ai GLM 4 32B — April 2025 base	2025-04 → 2025-04	5.18	GLM-4-32B-0414 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
ByteDance Seed Seed-OSS 36B — base	2025-08 → 2025-08	4.67	Seed-OSS-36B-Base : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Meta Chameleon 7B + 34B — pretraining	2024-06 → 2024-06	10.28	@2024-06-18 : Annonce primaire — source 1 Weights release June 18, 2024, not May paper publication.
Meta Code Llama — released 12-model increment	2023-08 → 2024-01	2.80	@2023-08-24 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2024-01-29 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
Meta CWM — pretraining + mid-training	2025-09 → 2025-09	4.49	@2025-09-24 : Annonce primaire — source 1 Release announcement September 24, 2025.
AI21 Labs Jamba 1.5 Large — base scenario	2024-08 → 2024-08	8.12	Jamba 1.5-Large : Publication Epoch (proxy) — source 1 · source 2 · source 3 · source 4 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
StepFun Step 3 — base scenario	2025-07 → 2025-07	7.39	@2025-07-01 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification. @2025-07-31 : Mois de travail — à vérifier — source 1 Curated working month for the budgeted release; open-weight availability still requires independent verification.
StepFun Step 3.5 Flash — base scenario	2026-02 → 2026-02	2.97	Step 3.5 Flash : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Ai2 OLMo 3 — 7B + 32B pretraining	2025-11 → 2025-11	2.35	Olmo 3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Mistral AI Mistral Small 4 — base scenario	2026-03 → 2026-03	1.75	Mistral Small 4 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Mistral AI Mistral Small 3 — 24B base scenario	2025-01 → 2025-01	4.67	Mistral Small 3 : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Mistral AI Mistral NeMo — 12B joint base	2024-07 → 2024-07	1.04	Mistral NeMo : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.
Mistral AI Mistral 7B — original base	2023-09 → 2023-09	0.19	@2023-09-27 : Annonce primaire — source 1 Release September 27, 2023, not October paper date.
Mistral AI Codestral 22B — one base scenario	2024-05 → 2024-05	1.92	Codestral : Publication Epoch (proxy) — source 1 · source 2 Exact named Epoch row; Publication date is a secondary proxy, not independently verified weight availability.

De 844 à 991 M$ : les familles manquantes

+148,80 M$ de nouveaux postes, puis −1,80 M$ de marge Cohere retirée pour éviter un recouvrement. Les hypothèses de génération et les prix n’ont pas été ajustés pour atteindre le milliard.

Passage du scénario V3 à V4 avec ajouts par organisation et retrait de la marge Cohere

Sans Cosmos : 955 M$. Sans les nouvelles bases dont l’indépendance reste hypothétique : 926 M$. Ces sensibilités conservent les autres hypothèses de V3. Le résultat soutient un ordre de grandeur, pas un minimum garanti.

Organisation / poste ajouté	Bas / central / haut (M$)	Hypothèses, sources et recouvrements
NVIDIA Cosmos 1 / Predict1 — allocation scenario	8.10 / 27.00 / 64.80	Scénario avec hypothèses Reported cluster 10,000 H100 over three months; 90-day approximation. ASSUME 25/50/75% of cluster-capacity hours attributable to retained model training (not utilisation measurement), at $1.5/2.5/4 per GPU-hour. Excludes separately adding constituent models. Allocation is not empirically calibrated. Entire released WFM group counted once; no separate Transfer1, tokenizer or inherited NeMo/T5 base added. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
NVIDIA Cosmos 3 Nano + Super — generator increment	4.32 / 9.72 / 43.20	Scénario avec hypothèses Architecture scenario: 6ND with one 8B/32B tower per token × (31.05+2.4)T/(17.86+1.9)T tokens. Token counts reported; effective training compute proxy uncertain. No use of ambiguous benchmark throughput. High doubles compute proxy for architecture/attention overhead. This is a sensitivity assumption, not a measured bound. Qwen3-VL upstream excluded; Edge/Alpamayo/posttraining excluded. Separate generator training only. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Cohere Command A — 111B base	0.00 / 21.58 / 79.92	Scénario avec hypothèses 6ND proxy; 111B active × [10, 18, 30]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Remove legacy Command R/R+ successor allowance first; Vision/Reasoning/Aya not added separately. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
Cohere Command A+ — 25B active base	0.00 / 6.75 / 24.00	Scénario avec hypothèses 6ND proxy; 25B active × [15, 25, 40]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Separate-generation assumption; no Vision/Translate/North increments added. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
xAI Grok 2 — 2024 base scenario	0.00 / 9.76 / 40.68	Scénario avec hypothèses 6ND proxy; 113B active × [4, 8, 15]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Grok1 separately counted; active113B is config-derived approximation; independent-run assumption; no additional instruct run. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
SK Telecom A.X K1	1.58 / 6.42 / 23.76	Scénario avec hypothèses 6ND proxy; 33B active × [10, 18, 30]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Card explicitly says from scratch; A.X3.1 omitted. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
SK Telecom A.X K2 — pretraining	1.30 / 2.92 / 6.49	Volumes renseignés, prix supposés 6ND proxy; 33B active × [8.2, 8.2, 8.2]T tokens; tokens reported; $0.8/1.8/4M per 1e24 FLOP. Card explicitly says from scratch; K1 not reused as a charged base. Posttraining not counted. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Alibaba Qwen 2 — dense 0.5/1.5/7/72B	2.73 / 6.15 / 13.67	Volumes renseignés, prix supposés Reported schedule: 0.5B×12T + (1.5+7+72)B×7T; 6ND; prices unchanged. 7T is a conservative reported slice, excludes long-context overhead. Pre-Qwen2.5 generation; each dense size once; excludes Qwen1/1.5 unresolved overlap. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Alibaba Qwen 2 57B-A14B — upcycling increment	0.30 / 0.68 / 1.51	Volumes renseignés, prix supposés 6ND proxy; 14B active × [4.5, 4.5, 4.5]T tokens; tokens reported; $0.8/1.8/4M per 1e24 FLOP. Only additional4.5T after Qwen2-7B; parent is already in preceding dense entry. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Z.ai GLM 4 32B — April 2025 base	2.30 / 5.18 / 11.52	Volumes renseignés, prix supposés 6ND proxy; 32B active × [15, 15, 15]T tokens; tokens reported; $0.8/1.8/4M per 1e24 FLOP. Distinct from GLM4.5 and5; Z1/Rumination/Flash/vision not added. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
ByteDance Seed Seed-OSS 36B — base	2.07 / 4.67 / 10.37	Volumes renseignés, prix supposés 6ND proxy; 36B active × [12, 12, 12]T tokens; tokens reported; $0.8/1.8/4M per 1e24 FLOP. One base only; woSyn alternative, Seed-Coder and BAGEL excluded. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Meta Chameleon 7B + 34B — pretraining	5.14 / 10.28 / 15.42	Volumes renseignés, prix supposés 5138888 GPU-hours × [1, 2, 3] USD/GPU-hour (assumed prices). Separate early-fusion runs; no Llama base charged. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Meta Code Llama — released 12-model increment	1.40 / 2.80 / 4.20	Volumes renseignés, prix supposés 1400000 GPU-hours × [1, 2, 3] USD/GPU-hour (assumed prices). Llama2 upstream already in V3; all12 includes70B. Older400k report is subset, not extra. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Meta CWM — pretraining + mid-training	2.00 / 4.49 / 9.98	Volumes renseignés, prix supposés 6ND proxy; 32B active × [13, 13, 13]T tokens; tokens reported; $0.8/1.8/4M per 1e24 FLOP. 8T+5T once across pretrain/SFT/final checkpoints; RL/SFT extra excluded. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
AI21 Labs Jamba 1.5 Large — base scenario	0.00 / 8.12 / 33.84	Scénario avec hypothèses 6ND proxy; 94B active × [4, 8, 15]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. One Large lineage;1.6/1.7/Mini/Jamba2 not separately costed. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
StepFun Step 3 — base scenario	0.00 / 7.39 / 27.36	Scénario avec hypothèses 6ND proxy; 38B active × [10, 18, 30]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. V3 Step5 explicitly excluded Step3; distinct-run assumption, vision extra excluded. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
StepFun Step 3.5 Flash — base scenario	0.00 / 2.97 / 10.56	Scénario avec hypothèses 6ND proxy; 11B active × [15, 25, 40]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Separate base assumed, not proven from architecture.3.7 and midtrain checkpoint not charged separately. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
Ai2 OLMo 3 — 7B + 32B pretraining	1.04 / 2.35 / 5.22	Volumes renseignés, prix supposés Reported5.93T/5.50T; nominal7B/32B;6ND. Later stages omitted. OLMo2 allowance limited to OLMo2; OLMo3 treated as new pretraining generation. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Mistral AI Mistral Small 4 — base scenario	0.00 / 1.75 / 6.24	Scénario avec hypothèses 6ND proxy; 6.5B active × [15, 25, 40]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Independent/upcycled status unknown. Leanstral not added. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
Mistral AI Mistral Small 3 — 24B base scenario	0.00 / 4.67 / 17.28	Scénario avec hypothèses 6ND proxy; 24B active × [10, 18, 30]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. One24B lineage;Small3.1/3.2,Magistral,DevstralSmall,VoxtralSmall not separately added. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire
Mistral AI Mistral NeMo — 12B joint base	0.23 / 1.04 / 4.32	Scénario avec hypothèses 6ND proxy; 12B active × [4, 8, 15]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Joint NVIDIA/Mistral model counted once under Mistral; NVIDIA mirror and Pixtral not additional base. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Mistral AI Mistral 7B — original base	0.05 / 0.19 / 0.69	Scénario avec hypothèses 6ND proxy; 7.2B active × [1.5, 2.5, 4]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. v0.2/v0.3/Mathstral not separately added; Mamba excluded. No actual dollar invoice; price and FLOP conversion assumptions. Source primaire
Mistral AI Codestral 22B — one base scenario	0.00 / 1.92 / 7.99	Scénario avec hypothèses 6ND proxy; 22.2B active × [4, 8, 15]T tokens; token volumes ASSUMED; $0.8/1.8/4M per 1e24 FLOP. Small2409 may share base: not added; Ministral8B2410 excluded. No actual dollar invoice; price and FLOP conversion assumptions. Independent-run scenario; reuse/upcycling not excluded. Source primaire

Un bas nul pour une base à filiation incertaine signifie « pas de nouvel entraînement de base ajouté dans cette branche », pas « coût de création nul ». 23 ajouts CSV · Méthode et sensibilités · Résumé JSON · Version précédente

Ajouts de la révision précédente

Mistral Large 3 : 7,97 M$ [1,97–29,52], avec volumes de tokens supposés. Nemotron Nano 2 : 2,66 M$ [1,18–5,91], avec 20T tokens publiés. Le second est sous le seuil de priorité de 30B, mais fait partie du périmètre global. Sources et calculs.

Chaque dépôt ≥30B, sans inventer sa filiation

Deux scénarios de calcul : préentraînement complet hypothétique ou entraînement supplémentaire seul. Les volumes publiés sont utilisés lorsqu’une composante individuelle a été reliée à sa source. La filiation reste à vérifier ; le coût du parent n’est pas inclus dans le second. Les valeurs génériques utilisent des volumes de tokens supposés. Quand les paramètres actifs sont inconnus, « équivalent dense » indique une approximation qui peut fortement surestimer un MoE.

Omissions identifiées, estimations à réconcilier

40 groupes candidats issus de la révision Claude ; liste historique. Le tableau V4 ci-dessus précise les postes désormais intégrés. Les estimations de cette liste peuvent se chevaucher et ne sont pas ajoutées automatiquement au total. Les budgets agrégés de familles ne sont pas des estimations individuelles.

NVIDIA Cosmos : scénarios désormais explicités

Les volumes de tokens sont publiés, mais les débits viennent d’un benchmark mixte image/vidéo. Leur interprétation et les nombres de GPU présentent des ambiguïtés. Le calcul par débit proposé pour Cosmos 3 reste écarté. V4 ajoute un proxy de calcul par tour pour Cosmos 3 et une allocation hypothétique de capacité pour Cosmos 1 ; ces deux montants sont des scénarios, pas des dépenses observées. Scénarios exploratoires et limites · Rapport primaire.

Ce que la révision change

Les tailles identiques ne prouvent plus une filiation ; les modèles BF16 et les dépôts déclarant un parent restent dans l’inventaire. Aucun budget de famille entier n’est attribué à un modèle choisi arbitrairement. Les cas inconnus ne deviennent pas gratuits.

Le registre reprend les 39 572 dépôts dont le snapshot du 17 octobre indique ≥30B paramètres, plus 81 candidats des catalogues et de l’audit. Les métadonnées manquantes, les poids réellement accessibles et les entraînements partagés restent à vérifier. Source du snapshot.

Combien de calcul partage-t-on gratuitement ?