[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mercury-3-diffusion-scaling":3,"topics-all":38,"news-related-713ebd28-7abc-4664-acd5-8ba7c6b50d97":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"713ebd28-7abc-4664-acd5-8ba7c6b50d97","扩散 LLM 能否 scale?Mercury 3 用 40 倍规模和 1M 上下文作答","Inception 发布扩散 LLM Mercury 3:规模较 Mercury 1 扩大逾 40 倍,上下文扩至 100 万 token,官方称每秒 1068 token、基准超 GPT-6 Luna 与 Claude Haiku 5.5;OpenRouter 实测中位约 254 token,落差明显。","扩散语言模型从诞生起就背着一个问题:并行生成的速度优势,能不能在规模放大后保住?过去一年多,自回归模型靠堆参数一路变强,扩散 LLM 却长期停留在\"小而快\"的证明阶段。Inception 昨天给出了自己的答案——第三代模型 Mercury 3,官方称之为迄今发布的最强扩散 LLM(官方博客:https:\u002F\u002Fwww.inceptionlabs.ai\u002Fblog\u002Fintroducing-mercury-3)。\n\n## 40 倍规模,速度没丢\n\nMercury 3 的核心叙事是\"text diffusion scales\"。官方数据显示,它的总参数量是 2025 年 6 月首发的 Mercury 1 的 40 倍以上,标准基准上的智能水平较 Mercury 1 mini 提升约 3 倍,但推理速度保持在每秒 1000 token 以上——NVIDIA GPU 上为 1068 token。上下文窗口从 Mercury 2.5 的 26 万 token 直接扩到 100 万。\n\n价格侧,每百万 token 输入 0.2 美元、输出 0.7 美元,发布期五折(0.1\u002F0.35 美元),缓存 token 按一折计费。\n\n## 官方基准与生产数据\n\n官方基准称 Mercury 3 在 GPQA Diamond、IFBench、SWE-Bench Atlas 等多项测试上超过 GPT-6 Luna 与 Claude Haiku 5.5,同时快 4-8 倍。生产侧的证据更具体:在 50 个真实 Mind2Web 导航任务上,Mercury 3 完成 50%,官方称明显高于 GPT-6.1-Sol,且 token 用量与步数相同、单 token 成本只有零头。智能眼镜搜索公司 Octen AI 给出的中位生成时间为 0.9 秒、端到端 2.2 秒。\n\n注意:上述跑分均为厂商自报,尚无独立复现。\n\n## 宣称与实测的裂缝\n\n第三方评测站 explainx 当天做了交叉核对:OpenRouter 单一 provider 的实测中位速度约每秒 254 token、P50 延迟 1.91 秒,与官方的 1068 token 相差逾 4 倍。另外两个信息官方没有给出:精确参数量未披露,权重也不开放——纯 API 交付(Inception API、OpenRouter、Baseten、Vercel,OpenAI 兼容接口,新账号送 1 亿 token)。\n\n## 为什么值得关心\n\n如果扩缩定律对扩散模型同样成立——这正是 Inception 的核心主张——那行业就多了一条独立于自回归的规模化路径:智能增长不必以推理变慢为代价。对浏览器 Agent、语音、实时搜索这类每一步都在等模型的场景,这是结构性利好;Inception 同场还推出首 token 低至 170 毫秒的 Mercury Voice 和对标 Jev 的 Mercury Decide,产品组合明显押注实时赛道。\n\n但 254 与 1068 的落差也提醒我们:厂商口中的\"每秒千 token\",和你自己网关里跑出来的,可能不是同一个数字。下次模型发布,先问一句\"实测呢\",再决定要不要把架构迁移提上日程。","https:\u002F\u002Fwww.inceptionlabs.ai\u002Fblog\u002Fintroducing-mercury-3","d34e492b-4ef4-4694-b6ad-fd4e58d0169a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"a4f801f4-4b85-43f7-acd5-85d7fc0d29a4","en","Mercury 3: Inception's Diffusion LLM Scales 40x With 1M Context","Inception's Mercury 3: 40x Mercury 1, 1M context, 1,068 tok\u002Fs claimed, $0.20\u002F$0.70 per M tokens. Third-party OpenRouter check: ~254 tok\u002Fs median.","Diffusion language models have carried one question since birth: can the speed advantage of parallel generation survive scaling? For over a year, autoregressive models grew stronger by stacking parameters, while diffusion LLMs stayed in the \"small but fast\" proof stage. Inception answered yesterday with its third-generation model Mercury 3, which the company calls the most capable diffusion LLM released to date (official blog: https:\u002F\u002Fwww.inceptionlabs.ai\u002Fblog\u002Fintroducing-mercury-3).\n\n## 40x Scale, Speed Intact\n\nMercury 3's core narrative is that text diffusion scales. Per the official announcement, its total parameter count is more than 40x that of Mercury 1 (launched June 2025), intelligence on standard benchmarks has risen roughly 3x from Mercury 1 mini, yet inference speed stays above 1,000 tokens per second — 1,068 tokens\u002Fsec on NVIDIA GPUs. The context window jumps from 260K tokens (Mercury 2.5) straight to 1 million.\n\nOn pricing: 0.20 USD per million input tokens and 0.70 USD per million output tokens, halved to 0.10\u002F0.35 USD during launch, with cached tokens billed at 90% off the input price.\n\n## Vendor Benchmarks and Production Data\n\nVendor benchmarks claim Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 on tests including GPQA Diamond, IFBench, and SWE-Bench Atlas while running 4-8x faster. The production evidence is more concrete: on 50 live Mind2Web navigation tasks, Mercury 3 completed 50%, which the company says is significantly more than GPT-6.1-Sol, using the same token volume and agent steps at a fraction of the per-token cost. Smart-glasses search startup Octen AI reported a median generation time of 0.9 seconds and end-to-end response of 2.2 seconds.\n\nNote: all scores above are vendor-reported and not yet independently reproduced.\n\n## The Gap Between Claim and Measurement\n\nThird-party review site explainx ran a cross-check the same day: a single OpenRouter provider showed a measured median of about 254 tokens\u002Fsec with 1.91s P50 latency — more than 4x below the official 1,068 figure. Two other details the company left out: the exact parameter count is undisclosed, and the weights are closed — API-only delivery (Inception API, OpenRouter, Baseten, Vercel, with an OpenAI-compatible interface and 100 million free tokens for new accounts).\n\n## Why It Matters\n\nIf scaling laws hold for diffusion — which is exactly Inception's central claim — the industry gains a second scaling path independent of autoregression: intelligence can grow without inference getting slower. For browser agents, voice, and real-time search, where every step waits on the model, that is a structural win; Inception also launched Mercury Voice (first answer token under 170ms) and Mercury Decide (benchmarking against Jev) in the same wave, a portfolio clearly betting on real-time AI.\n\nBut the 254-vs-1,068 gap is a reminder: the \"1,000+ tokens per second\" in a vendor's announcement and the number coming out of your own gateway may not be the same. Next time a model launches, ask \"what does it measure in practice?\" before putting an architecture migration on the roadmap.","mercury-3-diffusion-scaling","2026-10-19T23:09:33Z","2026-10-19T23:09:40.955062Z","2026-10-19T23:09:40.955075Z",true,"agent",142,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"178858d2-cb75-40e4-b3d1-6ac90add7d96","扩散模型上岗电话客服:推理开口只要320毫秒","mercury-voice-diffusion-llm","2026-09-29T23:07:20+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"8c6d408b-2913-48b2-bc1f-4e958ad854c5","Space Bunny 谜底揭晓:MiniMax M3.1 Flash,匿名期吞下 200 万亿 token","minimax-m3-1-flash-space-bunny","2026-10-19T21:10:12+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d6944f35-e3a6-4d5f-a4e1-a546836f7d75","微软基于 Qwen3.5-9B 推出决策模型 Microsoft-Decision-1","microsoft-decision-1-qwen-3-5-9b","2026-10-13T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"6ede1566-74c5-4bdd-9dd5-6d3f118db876","投机解码的重复陷阱:D-Loop 两轮循环破局","d-loop-diffusion-speculative-decoding","2026-10-06T15:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"8b50b8ec-0249-420f-a091-0d8b7efe6a18","Flash-dLLM 免训加速扩散语言模型:最高148倍","flash-dllm-io-aware-kv-cache","2026-09-23T15:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"b667e52f-ec7d-4ca4-8d9e-1db81e1a5616","DeepSeek论文:890字节KV缓存的三层架构账","deepseek-v41-flash-kv-cache-paper","2026-09-18T15:10:00+00:00"]