[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Ffr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards":3,"news-related-fr-\u002Ffr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards":412,"hreflang-deepseek-v4":431},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":391,"date":392,"description":393,"extension":394,"image":395,"meta":396,"navigation":397,"path":398,"seo":399,"seoDescription":400,"seoTitle":401,"stem":402,"tags":403,"translationKey":405,"updatedAt":8,"__hash__":411},"news_fr\u002Ffr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards.md","DeepSeek V4 : le modèle open-weight à 1 000 milliards de paramètres qui défie OpenAI","Paul Lefizelier",null,{"type":10,"value":11,"toc":381},"minimark",[12,29,34,48,63,79,83,98,114,117,121,124,130,145,156,171,175,182,203,206,210,213,315,321,327,333,336,340,373,375],[13,14,15,16,20,21,24,25,28],"p",{},"Ce 16 mars 2026, le lab chinois ",[17,18,19],"strong",{},"DeepSeek"," publie ",[17,22,23],{},"DeepSeek V4",", un modèle de langage à ",[17,26,27],{},"1 000 milliards de paramètres"," totaux — dont seulement 32 milliards actifs. Open-weight, téléchargeable et déployable en local, il rivalise avec les meilleurs modèles propriétaires du marché. Voici pourquoi les développeurs et builders IA doivent s'y intéresser.",[30,31,33],"h2",{"id":32},"quest-ce-que-deepseek-v4","Qu'est-ce que DeepSeek V4 ?",[13,35,36,37,40,41,43,44,47],{},"DeepSeek V4 repose sur une architecture ",[17,38,39],{},"MoE"," (Mixture of Experts). Le principe : le modèle contient ",[17,42,27],{}," au total, mais n'en active que ",[17,45,46],{},"32 milliards"," à chaque requête. Seuls les « experts » pertinents sont sollicités. Le résultat : des performances de modèle frontier avec une consommation de calcul bien inférieure.",[13,49,50,51,54,55,58,59,62],{},"C'est l'évolution directe de ",[17,52,53],{},"DeepSeek V3"," et ",[17,56,57],{},"DeepSeek R1",", les modèles qui avaient secoué la Silicon Valley début 2025. V4 pousse la logique MoE encore plus loin avec une fenêtre de contexte de ",[17,60,61],{},"128 000 tokens"," et un support multilingue renforcé — particulièrement en chinois et en anglais.",[13,64,65,66,69,70,78],{},"Le modèle est ",[17,67,68],{},"open-weight"," : ses poids sont ",[71,72,77],"a",{"href":73,"rel":74,"target":76},"https:\u002F\u002Fhuggingface.co\u002Fdeepseek-ai",[75],"nofollow","\\_blank","disponibles sur Hugging Face"," et peuvent être téléchargés, fine-tunés et déployés sur n'importe quelle infrastructure privée. Pas de dépendance à une API. Pas de lock-in.",[30,80,82],{"id":81},"des-performances-frontier-à-un-coût-fracassé","Des performances frontier à un coût fracassé",[13,84,85,86,89,90,93,94,97],{},"Sur les benchmarks de raisonnement, de code et de mathématiques, DeepSeek V4 se place au niveau de ",[17,87,88],{},"GPT-5.3"," (OpenAI), ",[17,91,92],{},"Gemini 3.1 Pro"," (Google) et ",[17,95,96],{},"Claude Sonnet 4.6"," (Anthropic). Les différences sont marginales selon les tâches.",[13,99,100,101,104,105,108,109,113],{},"La vraie rupture est économique. Le coût d'inférence de DeepSeek V4 est estimé à environ ",[17,102,103],{},"0,50 $ par million de tokens en entrée",". C'est ",[17,106,107],{},"10 fois moins"," que GPT-5.3 (~5 $) et nettement sous Claude Sonnet 4.6 (~3 $). Pour les builders qui orchestrent des ",[71,110,112],{"href":111},"\u002Ffr\u002Fblog\u002Fagents-ia-developpeurs-guide-outils-autonomes-2026","agents IA autonomes"," ou des pipelines RAG, cette différence de coût change l'équation.",[13,115,116],{},"Le coût d'entraînement lui-même reste une fraction de celui des labs américains. DeepSeek a hérité des optimisations radicales de R1 et V3 — des techniques d'architecture qui extraient le maximum de chaque puce disponible.",[30,118,120],{"id":119},"open-weight-ce-que-ça-change-concrètement-pour-les-builders","Open-weight : ce que ça change concrètement pour les builders",[13,122,123],{},"Un modèle open-weight (dont les poids sont librement téléchargeables) offre trois avantages décisifs par rapport aux modèles propriétaires accessibles uniquement via API.",[13,125,126,129],{},[17,127,128],{},"L'indépendance."," Pas de risque de changement de pricing, de rate limiting ou de coupure d'accès. Le modèle tourne sur votre infrastructure, sous votre contrôle.",[13,131,132,135,136,140,141,144],{},[17,133,134],{},"La personnalisation."," DeepSeek V4 peut être fine-tuné sur des données métier spécifiques. Les développeurs qui utilisent ",[71,137,139],{"href":138},"\u002Ffr\u002Fnews\u002Fvibecoding-outils-agentic-ai-dev-2026","Cursor, Replit ou d'autres IDE IA"," peuvent l'intégrer directement dans leurs workflows de ",[71,142,143],{"href":138},"vibecoding",".",[13,146,147,150,151,155],{},[17,148,149],{},"Le coût marginal."," En déploiement local ou sur cloud privé, le ",[71,152,154],{"href":153},"\u002Ffr\u002Fblog\u002Ftendances-tech-developpement-2026","coût d'inférence"," baisse encore — jusqu'à l'autohébergement gratuit pour les équipes qui disposent déjà de GPU.",[13,157,158,159,162,163,166,167,170],{},"Le retard des modèles open source sur les propriétaires s'est réduit à ",[17,160,161],{},"3 mois"," environ. Il y a un an, cet écart était de 12 à 18 mois. DeepSeek V4, aux côtés de ",[17,164,165],{},"Llama 4"," (Meta) et ",[17,168,169],{},"Mistral Large 3"," (Mistral AI), confirme cette accélération.",[30,172,174],{"id":173},"deepseek-et-la-géopolitique-de-lia","DeepSeek et la géopolitique de l'IA",[13,176,177,178,181],{},"DeepSeek V4 n'est pas qu'un modèle de langage. C'est un signal géopolitique. Le lab chinois démontre qu'il est possible d'entraîner un modèle frontier ",[17,179,180],{},"sans accès aux puces Nvidia H100",", dont l'exportation vers la Chine est restreinte par les États-Unis depuis 2022.",[13,183,184,185,189,190,193,194,197,198,202],{},"Les optimisations logicielles de ",[71,186,19],{"href":187,"rel":188,"target":76},"https:\u002F\u002Fwww.deepseek.com",[75]," — architecture MoE agressive, compression des gradients, parallélisme optimisé — compensent le déficit matériel. Pendant ce temps, ",[17,191,192],{},"ByteDance"," aurait acquis ",[17,195,196],{},"2,5 milliards de dollars"," de puces Nvidia via des intermédiaires en Asie du Sud-Est. La ",[71,199,201],{"href":200},"\u002Ffr\u002Fnews\u002Fexplosion-funding-ai-openai-anthropic-record-2026","course aux ressources compute"," entre Washington et Pékin s'intensifie.",[13,204,205],{},"Pour les régulateurs américains, DeepSeek V4 pose une question inconfortable : les restrictions d'export sur les puces suffisent-elles à freiner l'IA chinoise quand l'innovation logicielle contourne les barrières matérielles ?",[30,207,209],{"id":208},"faut-il-switcher-de-gpt-à-deepseek-v4","Faut-il switcher de GPT à DeepSeek V4 ?",[13,211,212],{},"Voici un comparatif honnête des principaux modèles frontier en mars 2026 :",[214,215,216,238],"table",{},[217,218,219],"thead",{},[220,221,222,226,229,232,235],"tr",{},[223,224,225],"th",{},"Modèle",[223,227,228],{},"Paramètres actifs",[223,230,231],{},"Open-weight",[223,233,234],{},"Contexte",[223,236,237],{},"Prix input\u002F1M tokens",[239,240,241,258,273,287,301],"tbody",{},[220,242,243,246,249,252,255],{},[244,245,23],"td",{},[244,247,248],{},"32 Md (MoE)",[244,250,251],{},"✅",[244,253,254],{},"128k",[244,256,257],{},"~0,5 $",[220,259,260,262,265,268,270],{},[244,261,88],{},[244,263,264],{},"N\u002FA",[244,266,267],{},"❌",[244,269,254],{},[244,271,272],{},"~5 $",[220,274,275,277,279,281,284],{},[244,276,92],{},[244,278,264],{},[244,280,267],{},[244,282,283],{},"1M",[244,285,286],{},"2 $",[220,288,289,291,293,295,298],{},[244,290,96],{},[244,292,264],{},[244,294,267],{},[244,296,297],{},"200k",[244,299,300],{},"3 $",[220,302,303,305,308,310,312],{},[244,304,165],{},[244,306,307],{},"~70 Md",[244,309,251],{},[244,311,254],{},[244,313,314],{},"gratuit",[13,316,317,320],{},[17,318,319],{},"DeepSeek V4 excelle"," pour les usages agents, le RAG (Retrieval-Augmented Generation), l'automation et les workflows où le volume de tokens est élevé. Son ratio performance\u002Fcoût est imbattable.",[13,322,323,326],{},[17,324,325],{},"Ses limites"," : l'écosystème d'outils est moins mature que celui d'OpenAI. La documentation est parfois en chinois uniquement. Et pour les entreprises soumises à des contraintes réglementaires strictes, utiliser un modèle entraîné en Chine peut poser des questions de conformité.",[13,328,329,332],{},[17,330,331],{},"Le verdict"," : DeepSeek V4 ne remplace pas GPT-5 ou Claude pour tous les cas d'usage. Mais il rend les modèles frontier accessibles à des équipes qui n'avaient pas le budget pour les API propriétaires.",[334,335],"hr",{},[30,337,339],{"id":338},"en-résumé","En résumé",[341,342,343,349,355,361,367],"ul",{},[344,345,346,348],"li",{},[17,347,23],{}," est un modèle IA open-weight de 1 000 milliards de paramètres totaux (32 milliards actifs), basé sur une architecture MoE (Mixture of Experts).",[344,350,351,354],{},[17,352,353],{},"Ses performances"," rivalisent avec GPT-5.3, Gemini 3.1 Pro et Claude Sonnet 4.6 sur les benchmarks de raisonnement, de code et de mathématiques.",[344,356,357,360],{},[17,358,359],{},"Son coût d'inférence"," est d'environ 0,50 $ par million de tokens en entrée — 10 fois moins que GPT-5.3 d'OpenAI.",[344,362,363,366],{},[17,364,365],{},"Le modèle est open-weight"," : téléchargeable sur Hugging Face, déployable en local, fine-tunable sans dépendance à une API propriétaire.",[344,368,369,372],{},[17,370,371],{},"DeepSeek V4 démontre"," que la Chine peut entraîner des modèles frontier malgré les restrictions américaines sur l'export des puces Nvidia H100.",[334,374],{},[13,376,377],{},[378,379,380],"em",{},"L'open-weight progresse à une vitesse qui doit inquiéter les labs propriétaires. Quand un modèle gratuit et téléchargeable rivalise avec des API à 5 $ le million de tokens, combien de temps le modèle commercial d'OpenAI ou d'Anthropic peut-il tenir sans évoluer ?",{"title":382,"searchDepth":383,"depth":383,"links":384},"",2,[385,386,387,388,389,390],{"id":32,"depth":383,"text":33},{"id":81,"depth":383,"text":82},{"id":119,"depth":383,"text":120},{"id":173,"depth":383,"text":174},{"id":208,"depth":383,"text":209},{"id":338,"depth":383,"text":339},"ai","2026-03-16T12:27:00","DeepSeek V4 est un modèle IA open-weight de 1 000 milliards de paramètres (32 Md actifs). Il rivalise avec GPT-5 et Claude à une fraction du coût.","md","\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards.jpg",{},true,"\u002Ffr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards",{"title":6,"description":393},"DeepSeek V4 : 1 000 milliards de paramètres, open-weight, performances frontier. Architecture MoE, benchmarks, coûts et impact pour les développeurs IA.","DeepSeek V4 : le modèle open-weight qui défie GPT-5","fr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards",[404,405,68,406,407,408,409,410],"deepseek","deepseek-v4","llm","modele-ia","chine","open-source","frontier-model","xPAAD3JQWXDG3jokII50jJ6vKf0nZJaDvxo0-Es5iQQ",[413,419,425],{"title":414,"description":415,"path":416,"date":417,"image":418,"category":143},"Les cloud agents Cursor construisent, hébergent et publient des apps live","Les cloud agents Cursor peuvent scaffolder une app web, provisionner Postgres et le SSO Cursor, lancer des contrôles de sécurité, puis publier une URL partageable — sans compte d'hébergement séparé.","\u002Ffr\u002Fnews\u002Fcursor-live-apps-agents-cloud-deploiement-postgres-auth","2026-11-07T08:00:00","\u002Fimages\u002Fnews\u002Fcursor-live-apps-cloud-agents-deploy-postgres-auth.jpg",{"title":420,"description":421,"path":422,"date":423,"image":424,"category":391},"Google rend Gemini Omni 1.2 Flash généralement disponible, avec des vidéos de 15 secondes","Google DeepMind baisse le prix d'Omni Flash de 30 %, à 0,07 $ par seconde en 720p, passe de 10 à 15 secondes de clip, et ouvre le modèle sur l'API Gemini et Google Cloud.","\u002Ffr\u002Fnews\u002Fgoogle-gemini-omni-1-2-flash-video-15-secondes","2026-11-06T08:00:00","\u002Fimages\u002Fnews\u002Fgoogle-gemini-omni-1-2-flash-15-second-video.jpg",{"title":426,"description":427,"path":428,"date":429,"image":430,"category":391},"Anthropic aurait signé son premier accord de licence avec un éditeur, l'AP","Selon Adweek, Anthropic a conclu un accord de licence de contenus et de données avec l'Associated Press juste avant les élections de mi-mandat américaines, son premier avec un éditeur numérique. Aucune des deux entreprises ne l'a confirmé officiellement, et les conditions ne sont pas publiques.","\u002Ffr\u002Fnews\u002Fanthropic-premier-accord-licence-editeurs-associated-press","2026-11-05T12:35:00","\u002Fimages\u002Fnews\u002Fanthropic-ap-publisher-licensing-deal.jpg",{"en":432,"fr":788,"es":8},{"id":433,"title":434,"author":7,"author_avatar":8,"body":435,"category":391,"date":392,"description":777,"extension":394,"image":395,"meta":778,"navigation":397,"path":779,"seo":780,"seoDescription":781,"seoTitle":782,"stem":783,"tags":784,"translationKey":405,"updatedAt":8,"__hash__":787},"news_en\u002Fnews\u002Fdeepseek-v4-open-weight-ai-model-1-trillion-parameters.md","DeepSeek V4: The 1-Trillion Parameter Open-Weight Model Challenging OpenAI",{"type":10,"value":436,"toc":769},[437,450,454,468,481,492,496,507,528,531,535,538,544,557,567,580,584,591,608,611,615,618,707,713,719,725,727,731,762,764],[13,438,439,440,442,443,445,446,449],{},"On March 16, 2026, Chinese lab ",[17,441,19],{}," releases ",[17,444,23],{},", a language model with ",[17,447,448],{},"1 trillion total parameters"," — of which only 32 billion are active per query. Open-weight, downloadable and locally deployable, it rivals the best proprietary models on the market. Here's why AI developers and builders should pay attention.",[30,451,453],{"id":452},"what-is-deepseek-v4","What Is DeepSeek V4?",[13,455,456,457,459,460,463,464,467],{},"DeepSeek V4 is built on a ",[17,458,39],{}," (Mixture of Experts) architecture. The concept: the model contains ",[17,461,462],{},"1 trillion parameters"," in total but only activates ",[17,465,466],{},"32 billion"," per request. Only the relevant \"experts\" are called upon. The result: frontier-level performance with far less compute.",[13,469,470,471,473,474,476,477,480],{},"It's the direct evolution of ",[17,472,53],{}," and ",[17,475,57],{},", the models that shook Silicon Valley in early 2025. V4 pushes the MoE approach further with a ",[17,478,479],{},"128,000-token context window"," and enhanced multilingual support — particularly in Chinese and English.",[13,482,483,484,486,487,491],{},"The model is ",[17,485,68],{},": its weights are ",[71,488,490],{"href":73,"rel":489,"target":76},[75],"available on Hugging Face"," and can be downloaded, fine-tuned and deployed on any private infrastructure. No API dependency. No vendor lock-in.",[30,493,495],{"id":494},"frontier-performance-at-a-fraction-of-the-cost","Frontier Performance at a Fraction of the Cost",[13,497,498,499,89,501,503,504,506],{},"On reasoning, code and math benchmarks, DeepSeek V4 performs on par with ",[17,500,88],{},[17,502,92],{}," (Google) and ",[17,505,96],{}," (Anthropic). Differences are marginal depending on the task.",[13,508,509,510,513,514,517,518,522,523,527],{},"The real breakthrough is economic. DeepSeek V4's inference cost sits at roughly ",[17,511,512],{},"$0.50 per million input tokens",". That's ",[17,515,516],{},"10x cheaper"," than GPT-5.3 (",[519,520,521],"del",{},"$5) and well below Claude Sonnet 4.6 (","$3). For builders orchestrating ",[524,525,526],"span",{},"INTERNAL LINK: autonomous AI agents | AI agents in enterprise article"," or RAG pipelines, this cost difference changes the equation.",[13,529,530],{},"Training costs themselves remain a fraction of what US labs spend. DeepSeek inherited the radical optimizations from R1 and V3 — architectural techniques that squeeze maximum performance from every available chip.",[30,532,534],{"id":533},"open-weight-what-it-concretely-changes-for-builders","Open-Weight: What It Concretely Changes for Builders",[13,536,537],{},"An open-weight model (one whose weights are freely downloadable) offers three decisive advantages over proprietary models accessible only through APIs.",[13,539,540,543],{},[17,541,542],{},"Independence."," No risk of pricing changes, rate limiting or access cuts. The model runs on your infrastructure, under your control.",[13,545,546,549,550,553,554,144],{},[17,547,548],{},"Customization."," DeepSeek V4 can be fine-tuned on domain-specific data. Developers using ",[524,551,552],{},"INTERNAL LINK: Cursor, Replit or other AI IDEs | vibe coding article"," can integrate it directly into their ",[524,555,556],{},"INTERNAL LINK: vibe coding workflows | vibe coding article on Idlen",[13,558,559,562,563,566],{},[17,560,561],{},"Marginal cost."," On local or private cloud deployments, ",[524,564,565],{},"INTERNAL LINK: inference costs | LLM price drop article"," drop further — down to free self-hosting for teams that already have GPUs.",[13,568,569,570,573,574,576,577,579],{},"The gap between open-source and proprietary models has shrunk to roughly ",[17,571,572],{},"3 months",". A year ago, that gap was 12 to 18 months. DeepSeek V4, alongside ",[17,575,165],{}," (Meta) and ",[17,578,169],{}," (Mistral AI), confirms this acceleration.",[30,581,583],{"id":582},"deepseek-and-the-geopolitics-of-ai","DeepSeek and the Geopolitics of AI",[13,585,586,587,590],{},"DeepSeek V4 is more than a language model. It's a geopolitical signal. The Chinese lab demonstrates that training a frontier model ",[17,588,589],{},"without access to Nvidia H100 chips"," — whose export to China has been restricted by the US since 2022 — is possible.",[13,592,593,596,597,599,600,603,604,607],{},[71,594,19],{"href":187,"rel":595,"target":76},[75],"'s software optimizations — aggressive MoE architecture, gradient compression, optimized parallelism — compensate for the hardware deficit. Meanwhile, ",[17,598,192],{}," reportedly acquired ",[17,601,602],{},"$2.5 billion"," worth of Nvidia chips through intermediaries in Southeast Asia. The ",[524,605,606],{},"INTERNAL LINK: compute resource race | AI geopolitics article"," between Washington and Beijing is intensifying.",[13,609,610],{},"For US regulators, DeepSeek V4 raises an uncomfortable question: are chip export restrictions enough to slow Chinese AI when software innovation routes around hardware barriers?",[30,612,614],{"id":613},"should-you-switch-from-gpt-to-deepseek-v4","Should You Switch From GPT to DeepSeek V4?",[13,616,617],{},"Here's an honest comparison of the main frontier models in March 2026:",[214,619,620,638],{},[217,621,622],{},[220,623,624,627,630,632,635],{},[223,625,626],{},"Model",[223,628,629],{},"Active parameters",[223,631,231],{},[223,633,634],{},"Context",[223,636,637],{},"Price input\u002F1M tokens",[239,639,640,654,667,680,693],{},[220,641,642,644,647,649,651],{},[244,643,23],{},[244,645,646],{},"32B (MoE)",[244,648,251],{},[244,650,254],{},[244,652,653],{},"~$0.50",[220,655,656,658,660,662,664],{},[244,657,88],{},[244,659,264],{},[244,661,267],{},[244,663,254],{},[244,665,666],{},"~$5",[220,668,669,671,673,675,677],{},[244,670,92],{},[244,672,264],{},[244,674,267],{},[244,676,283],{},[244,678,679],{},"$2",[220,681,682,684,686,688,690],{},[244,683,96],{},[244,685,264],{},[244,687,267],{},[244,689,297],{},[244,691,692],{},"$3",[220,694,695,697,700,702,704],{},[244,696,165],{},[244,698,699],{},"~70B",[244,701,251],{},[244,703,254],{},[244,705,706],{},"free",[13,708,709,712],{},[17,710,711],{},"DeepSeek V4 excels"," at agent workflows, RAG (Retrieval-Augmented Generation), automation and high-token-volume use cases. Its performance-to-cost ratio is unmatched.",[13,714,715,718],{},[17,716,717],{},"Its limitations",": the tooling ecosystem is less mature than OpenAI's. Documentation is sometimes Chinese-only. And for enterprises with strict regulatory requirements, using a model trained in China may raise compliance concerns.",[13,720,721,724],{},[17,722,723],{},"The verdict",": DeepSeek V4 doesn't replace GPT-5 or Claude for every use case. But it makes frontier models accessible to teams that couldn't afford proprietary API pricing.",[334,726],{},[30,728,730],{"id":729},"key-takeaways","Key Takeaways",[341,732,733,738,744,750,756],{},[344,734,735,737],{},[17,736,23],{}," is a 1-trillion parameter open-weight AI model (32 billion active) built on a MoE (Mixture of Experts) architecture.",[344,739,740,743],{},[17,741,742],{},"Its performance"," rivals GPT-5.3, Gemini 3.1 Pro and Claude Sonnet 4.6 on reasoning, code and math benchmarks.",[344,745,746,749],{},[17,747,748],{},"Its inference cost"," is roughly $0.50 per million input tokens — 10x cheaper than OpenAI's GPT-5.3.",[344,751,752,755],{},[17,753,754],{},"The model is open-weight",": downloadable on Hugging Face, locally deployable, fine-tunable with no proprietary API dependency.",[344,757,758,761],{},[17,759,760],{},"DeepSeek V4 proves"," that China can train frontier models despite US restrictions on Nvidia H100 chip exports.",[334,763],{},[13,765,766],{},[378,767,768],{},"Open-weight is advancing at a pace that should worry proprietary labs. When a free, downloadable model rivals APIs priced at $5 per million tokens, how long can the business model of OpenAI or Anthropic hold without evolving?",{"title":382,"searchDepth":383,"depth":383,"links":770},[771,772,773,774,775,776],{"id":452,"depth":383,"text":453},{"id":494,"depth":383,"text":495},{"id":533,"depth":383,"text":534},{"id":582,"depth":383,"text":583},{"id":613,"depth":383,"text":614},{"id":729,"depth":383,"text":730},"DeepSeek V4 is a 1-trillion parameter open-weight AI model (32B active). It rivals GPT-5 and Claude at a fraction of the cost. Full breakdown.",{},"\u002Fnews\u002Fdeepseek-v4-open-weight-ai-model-1-trillion-parameters",{"title":434,"description":777},"DeepSeek V4: 1 trillion parameters, open-weight, frontier-level performance. MoE architecture, benchmarks, costs and impact for AI developers.","DeepSeek V4: the open-weight model challenging GPT-5","news\u002Fdeepseek-v4-open-weight-ai-model-1-trillion-parameters",[404,405,68,406,785,786,409,410],"ai-model","china","g9uTymE5kFRXnu47WgZtt2RGzYWqnwJAObiVh2AAS3c",{"id":5,"title":6,"author":7,"author_avatar":8,"body":789,"category":391,"date":392,"description":393,"extension":394,"image":395,"meta":1027,"navigation":397,"path":398,"seo":1028,"seoDescription":400,"seoTitle":401,"stem":402,"tags":1029,"translationKey":405,"updatedAt":8,"__hash__":411},{"type":10,"value":790,"toc":1019},[791,799,801,809,817,824,826,834,842,844,846,848,852,860,866,874,876,880,891,893,895,897,975,979,983,987,989,991,1013,1015],[13,792,15,793,20,795,24,797,28],{},[17,794,19],{},[17,796,23],{},[17,798,27],{},[30,800,33],{"id":32},[13,802,36,803,40,805,43,807,47],{},[17,804,39],{},[17,806,27],{},[17,808,46],{},[13,810,50,811,54,813,58,815,62],{},[17,812,53],{},[17,814,57],{},[17,816,61],{},[13,818,65,819,69,821,78],{},[17,820,68],{},[71,822,77],{"href":73,"rel":823,"target":76},[75],[30,825,82],{"id":81},[13,827,85,828,89,830,93,832,97],{},[17,829,88],{},[17,831,92],{},[17,833,96],{},[13,835,100,836,104,838,108,840,113],{},[17,837,103],{},[17,839,107],{},[71,841,112],{"href":111},[13,843,116],{},[30,845,120],{"id":119},[13,847,123],{},[13,849,850,129],{},[17,851,128],{},[13,853,854,135,856,140,858,144],{},[17,855,134],{},[71,857,139],{"href":138},[71,859,143],{"href":138},[13,861,862,150,864,155],{},[17,863,149],{},[71,865,154],{"href":153},[13,867,158,868,162,870,166,872,170],{},[17,869,161],{},[17,871,165],{},[17,873,169],{},[30,875,174],{"id":173},[13,877,177,878,181],{},[17,879,180],{},[13,881,184,882,189,885,193,887,197,889,202],{},[71,883,19],{"href":187,"rel":884,"target":76},[75],[17,886,192],{},[17,888,196],{},[71,890,201],{"href":200},[13,892,205],{},[30,894,209],{"id":208},[13,896,212],{},[214,898,899,913],{},[217,900,901],{},[220,902,903,905,907,909,911],{},[223,904,225],{},[223,906,228],{},[223,908,231],{},[223,910,234],{},[223,912,237],{},[239,914,915,927,939,951,963],{},[220,916,917,919,921,923,925],{},[244,918,23],{},[244,920,248],{},[244,922,251],{},[244,924,254],{},[244,926,257],{},[220,928,929,931,933,935,937],{},[244,930,88],{},[244,932,264],{},[244,934,267],{},[244,936,254],{},[244,938,272],{},[220,940,941,943,945,947,949],{},[244,942,92],{},[244,944,264],{},[244,946,267],{},[244,948,283],{},[244,950,286],{},[220,952,953,955,957,959,961],{},[244,954,96],{},[244,956,264],{},[244,958,267],{},[244,960,297],{},[244,962,300],{},[220,964,965,967,969,971,973],{},[244,966,165],{},[244,968,307],{},[244,970,251],{},[244,972,254],{},[244,974,314],{},[13,976,977,320],{},[17,978,319],{},[13,980,981,326],{},[17,982,325],{},[13,984,985,332],{},[17,986,331],{},[334,988],{},[30,990,339],{"id":338},[341,992,993,997,1001,1005,1009],{},[344,994,995,348],{},[17,996,23],{},[344,998,999,354],{},[17,1000,353],{},[344,1002,1003,360],{},[17,1004,359],{},[344,1006,1007,366],{},[17,1008,365],{},[344,1010,1011,372],{},[17,1012,371],{},[334,1014],{},[13,1016,1017],{},[378,1018,380],{},{"title":382,"searchDepth":383,"depth":383,"links":1020},[1021,1022,1023,1024,1025,1026],{"id":32,"depth":383,"text":33},{"id":81,"depth":383,"text":82},{"id":119,"depth":383,"text":120},{"id":173,"depth":383,"text":174},{"id":208,"depth":383,"text":209},{"id":338,"depth":383,"text":339},{},{"title":6,"description":393},[404,405,68,406,407,408,409,410]]