[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-quasar-438b-multiverse-europe-top-model":3,"topics-all":35,"news-related-3fba3833-d62f-49fb-a59e-a78630701194":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"3fba3833-d62f-49fb-a59e-a78630701194","Quasar 438B 拿下欧洲最高分:43 分、15.3 秒挤进全球第一梯队","Multiverse Computing 发布首个大模型 Quasar 438B:AA 智能指数 43 分,官方称欧洲模型最高;500 token 回复 15.3 秒,但定价与 token 冗长度均明显高于同档中位。","欧洲 AI 第一次摸到了全球第一梯队的门槛。Multiverse Computing 发布了公司历史上第一个大模型 Quasar 438B——一个 4380 亿参数的推理模型,主打企业级 Agent 和编码场景,支持英语和西班牙语。在 Artificial Analysis 智能指数 v4.1.1 上,它拿到 43 分,官方称之为欧洲模型的最高成绩;在 AA 收录的全部 178 个模型里,这个分数排到第 13 位。\n\n## 43 分,放在全球是什么位置\n\nv4.1.1 指数由九项评测组成:GDPval-AA v2、τ³-Banking、Terminal-Bench v2.1、SciCode、Humanity's Last Exam、GPQA Diamond、CritPt、AA-Omniscience 和 AA-LCR。\n\n放在欧洲同行里看,43 分比 Mistral Medium 3.5 的 30 分高出 13 分;比 NVIDIA Nemotron 3 Ultra 的 38 分高 5 分——而后者参数量还多出 1120 亿;也比 Inkling 的 42 分高 1 分。梯队顶端是 Claude Opus 5 的 63 分,Quasar 与前沿的差距仍然清晰,但\"欧洲第一\"的位置有了数据支撑。AA 自己的口径也印证了分数含金量:同档推理模型的中位数只有 18 分。\n\n## 速度是第二个卖点\n\nQuasar 输出 500 个 token(含思考时间)只要 15.3 秒。官方对比中只有三个模型更快:Nemotron 3.5 Lightning(9.4 秒,指数 24)、Gemini 3.5 Flash-Lite(10.8 秒,指数 37)、Gemini 3.7 Flash(11.5 秒,指数 56)——真正又快又强的只有最后一个。反过来看,Mistral Medium 3.5 跑同样的回复要 18.8 秒,Inkling 要 48.3 秒,超过 Quasar 的两倍。\n\n长上下文是它最接近前沿组的地方:AA-LCR 拿到 75.0,与 Grok 4.6 (high) 持平,距 Claude Opus 5 的 75.7 只差 0.7,也只落后 Qwen3.8 2.4T A95B 的 75.3 零点几分。短板同样明显——Terminal-Bench v2.1 的 69.3 分落后 Opus 5 的 89.1 分近 20 分,官方也承认这是余量最大的一项,下一轮工作会优先补这里。\n\n## 独立数据揭示了另一面\n\nArtificial Analysis 的独立评测页给出了官方通稿没强调的数字。输出速度 176.2 token\u002F秒(中位数 105.2),首 token 延迟 1.06 秒(中位数 2.05 秒),1M token 上下文,2026 年 8 月 10 日上线——性能确实过关。\n\n但价格和冗长度都站在贵的一侧:输入 $0.60\u002F百万 token、输出 $1.80\u002F百万 token,分别是同档中位数($0.25\u002F$0.90)的 2.4 倍和 2 倍。更扎眼的是 token 用量:跑完整个智能指数,Quasar 输出了 3.5 亿个 token,是中位数 6700 万的五倍多;AA 为此花了 $1047.71 的评测成本。推理模型的\"长思考\"在它身上格外突出。另外,它是专有模型,权重不开放,只能通过 API 使用。\n\n## 怎么看\n\nMultiverse Computing 此前的立足点是\"让 AI 更高效、更可部署\",Quasar 是这套方法论第一次被推进 400B+ 参数量级:多步任务要规划、用工具、跑代码、扛长上下文,同时把延迟压到能塞进交互式产品而非批处理作业的水平。模型通过自家的 CompactifAI API 提供,团队不用先搭基础设施就能直接测试。\n\n对欧洲来说,这至少证明了一件事:在 Mistral 之后,欧洲模型第一次有了进入全球前 15 的选项。但 Quasar 的账本也提醒买家——分数只是入场券,按 token 计费的世界里,冗长度直接决定月底账单。43 分很好,3.5 亿 token 的思考量,才是采购时要算的第二笔账。\n\n原文: https:\u002F\u002Fmultiversecomputing.com\u002Fresources\u002Fintroducing-quasar-438b-europe-s-leading-ai-model ;独立数据: https:\u002F\u002Fartificialanalysis.ai\u002Fmodels\u002Fquasar-438b","https:\u002F\u002Fmultiversecomputing.com\u002Fresources\u002Fintroducing-quasar-438b-europe-s-leading-ai-model","e22390fb-4d2e-4b0d-8777-3f9af3c151c9",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"da4dfbf9-2388-496e-a7a8-320760dbfc86","en","Quasar 438B: Europe's Top-Scoring Model Enters the Global Frontier at 43","Multiverse Computing has released Quasar 438B, its first large model: 43 on the Artificial Analysis Intelligence Index (a European best, per the company), 15.3s per 500 tokens — but pricing and token verbosity run well above the class median.","European AI has touched the threshold of the global frontier for the first time. Multiverse Computing has released Quasar 438B, the first large model in the company's history — a 438-billion-parameter reasoning model aimed at enterprise-scale agents and coding, supporting English and Spanish. It scores 43 on the Artificial Analysis Intelligence Index v4.1.1, which the company calls the highest result of any European model; across all 178 models tracked by AA, that score ranks 13th.\n\n## Where 43 Points Stands Globally\n\nThe v4.1.1 index combines nine evaluations: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR.\n\nAmong European peers, 43 is 13 points ahead of Mistral Medium 3.5 at 30; 5 ahead of NVIDIA Nemotron 3 Ultra at 38 — which carries 112 billion more parameters; and 1 ahead of Inkling at 42. The frontier is led by Claude Opus 5 at 63, so the gap remains clear, but the \"Europe's best\" claim now has data behind it. AA's own figures back the score's weight: the median among comparable reasoning models is just 18.\n\n## Speed Is the Second Selling Point\n\nQuasar outputs 500 tokens (thinking time included) in 15.3 seconds. Only three models in the official comparison are faster: Nemotron 3.5 Lightning (9.4s, index 24), Gemini 3.5 Flash-Lite (10.8s, index 37), and Gemini 3.7 Flash (11.5s, index 56) — only the last of those is both faster and more capable. Read the other way, Mistral Medium 3.5 needs 18.8 seconds for the same response and Inkling needs 48.3, more than double Quasar's time.\n\nLong context is where it comes closest to the frontier group: 75.0 on AA-LCR, level with Grok 4.6 (high), 0.7 behind Claude Opus 5 at 75.7 and just behind Qwen3.8 2.4T A95B at 75.3. The weak spot is equally clear — 69.3 on Terminal-Bench v2.1 trails Opus 5's 89.1 by nearly 20 points, and the company acknowledges this is the evaluation with the most headroom, where the next round of work is aimed.\n\n## The Independent Data Tells Another Story\n\nArtificial Analysis' independent model page provides numbers the announcement didn't emphasize. Output speed is 176.2 tokens\u002Fs (median: 105.2), time to first token is 1.06s (median: 2.05s), context window is 1M tokens, and the model went live on August 10, 2026 — the performance genuinely holds up.\n\nBut both price and verbosity sit on the expensive side: $0.60 per 1M input tokens and $1.80 per 1M output tokens — 2.4x and 2x the class medians ($0.25\u002F$0.90). More striking is token use: running the full Intelligence Index, Quasar generated 350 million output tokens, more than five times the median of 67M; AA spent $1,047.71 evaluating it. The \"long thinking\" of reasoning models is especially pronounced here. It is also a proprietary model — weights are not available, API access only.\n\n## What to Make of It\n\nMultiverse Computing built its position on making AI more efficient and deployable, and Quasar pushes that methodology into the 400B+ parameter class for the first time: multi-step tasks that need planning, tool use, code execution and large context, with latency kept low enough to sit inside an interactive product rather than a batch job. The model is served through the company's CompactifAI API, so teams can test it without standing up infrastructure.\n\nFor Europe, this at least proves one thing: after Mistral, European models now have their first option inside the global top 15. But Quasar's ledger is also a reminder to buyers — scores are just the entry ticket, and in a per-token billing world, verbosity directly decides the monthly invoice. 43 points is good; the 350-million-token thinking bill is the second calculation to run before procurement.\n\nSource: https:\u002F\u002Fmultiversecomputing.com\u002Fresources\u002Fintroducing-quasar-438b-europe-s-leading-ai-model ; independent data: https:\u002F\u002Fartificialanalysis.ai\u002Fmodels\u002Fquasar-438b","quasar-438b-multiverse-europe-top-model","2026-09-02T15:10:00Z","2026-09-02T15:12:05.728522Z","2026-09-02T15:12:05.728531Z",true,"agent",476,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"2c7a8ca0-ab46-454b-ae16-aa7e7b413753","Ultramarine 1 Flash:不可验证任务RL","ultramarine-1-flash-non-verifiable-rl","2026-10-20T07:07:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"5a5efb41-7037-46cb-b93e-6842f9ff2d62","UltraBERT:91.0 GLUE+47.7 BEIR,BERT 配方重审后 encoder 双榜登顶","ultrabert-encoder-dual-lm-heads","2026-10-14T17:10:59+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"27240196-3ac0-4af3-aec7-b0a4d3e7c990","智谱GLM-Legal法律Agent模型:Harvey榜前十最便宜","glm-legal-harvey-lab-cheapest","2026-10-13T15:20:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"6cd8c9b7-b493-4ad7-bd20-a907b9fb70f4","Claude 5.5 家族首发:Opus 5.5 缓存读取降六成,典型负载省四成","claude-opus-5-5-cost-cut-benchmarks","2026-09-22T21:10:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"b0587a16-6bf9-46cc-932b-23d7f37fed30","Grok 4.7 发布:定价持平跑分大涨,复测现分歧","grok-4-7-benchmarks-pricing","2026-09-21T23:09:13+00:00"]