[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mercury-voice-diffusion-llm":3,"topics-all":38,"news-related-178858d2-cb75-40e4-b3d1-6ac90add7d96":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},"178858d2-cb75-40e4-b3d1-6ac90add7d96","扩散模型上岗电话客服:推理开口只要320毫秒","Inception Labs 把扩散大模型塞进语音助手:Mercury Voice 正式商用,首个答案 token 中位 320 毫秒,官方口径比 GPT-6 Luna 快 5.9 倍,自报复合基准超过 GLM-5.3-Flash,支持推理与工具调用,128K 上下文,每分钟对话约 0.009 美元。","语音赛道有个不成文的生理学约束:人说完话,机器得在 500 毫秒内接上话头,再慢,对话就开始尴尬。为了守住这条红线,业界默认的解法是放弃推理——Gemma 4 31B、GPT-4.1 这类不带思考的模型统治语音 Agent,不是因为它们聪明,而是因为会推理的模型开口太慢。Inception Labs 9 月 29 日面向企业客户正式商用的 Mercury Voice,想用扩散架构把这道选择题拆掉——它两周前刚随 Mercury 2.5 预览过。\n\n## 320 毫秒是怎么来的\n\nMercury Voice 是专为语音 Agent 调优的扩散大模型(dLLM)。官方博客主打「首个答案 token 时间」(TTFAT):在真实客服语料上中位数 320 毫秒、p95 仅 750 毫秒,官方口径比不开推理的 GPT-6 Luna 快 5.9 倍;低推理档下,它是对比阵容里唯一中位数低于 500 毫秒对话预算的模型,p95 也快过除 Cerebras 上 GPT-OSS-120B 之外所有对手的中位数。尾部延迟被专门强调:p95 若是四秒,约二十轮对话就翻车一轮。底牌是扩散模型的老本行——并行解码替代逐 token 生成,Mercury 家族此前在文本侧跑出过每秒千 token、在搜索场景号称单查询能跑一百遍,这次是把同一机制搬进语音管线的推理槽位。\n\n## 会推理,也能打电话\n\n按官方自报,Mercury Voice 在 τ³-bench Telecom\u002FRetail\u002FAirline、IFBench 与 BFCL v4 组成的复合基准上超过 GPT-6 Luna、GLM-5.3-Flash、Gemini 3.5 Flash-Lite 和 Qwen3.5-397B。「推理模型塞进语音延迟预算」本身就是新信息:它提供低中高三档推理力度,128K 上下文、最高 50K 输出 token,接口兼容 OpenAI API,可直接插进 LiveKit、Pipecat、Vapi、Retell 等语音框架的 LLM 槽位。\n\n## 一分钟 0.9 美分的账\n\n定价每百万输入 token 0.40 美元、输出 1.50 美元,上线期五折(0.20\u002F0.75 美元)。按官方给的典型语音负载折算约每分钟对话 0.009 美元,对标 GPT-4.1 的 0.045 美元\u002F分钟,便宜约 5 倍。落地案例:Audivi AI 用它做全自动得来速点单;Altur 在金融机构的还款谈判通话里测出「最快」;OpenCall 生产环境中位模型延迟约 170 毫秒。引语都出自官方博客,但至少是真实付费客户。\n\n## 冷水与看点\n\n基准全部自报、第三方评测缺位,是这类发布的常规保留节目。更值得看的是轨迹:扩散 LLM 从文本到搜索、再到实时语音推理,每一步都咬住「既要低延迟、又要会思考」的场景。语音的延迟红线由人类生理钉死在 500 毫秒,而推理模型的思考还停在秒级——并行解码可能是目前唯一同时满足两边的工程答案。下次你打电话没听出对面是 AI,别只夸拟人,想想它是不是把「想」和「说」压进了同一瞬间。\n\n参考:https:\u002F\u002Fwww.inceptionlabs.ai\u002Fblog\u002Fintroducing-mercury-voice\n","https:\u002F\u002Fwww.inceptionlabs.ai\u002Fblog\u002Fintroducing-mercury-voice","a060ea51-5aec-4dfc-8c7a-898191593948",[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},"2113ae3a-d5c9-47b4-b3d0-8b3e10da147f","en","Mercury Voice: a diffusion LLM for real-time voice agents","Inception Labs ships Mercury Voice, a diffusion LLM for voice agents: 320 ms median first answer token, 5.9x faster than GPT-6 Luna.","Voice agents live by an unspoken physiological rule: answer within about 500 milliseconds after the caller finishes, or the pause starts to feel awkward. To hold that line, the industry has mostly given up reasoning — non-reasoning models such as Gemma 4 31B and GPT-4.1 are the default inside voice stacks not because they are smart, but because reasoning models take too long to open their mouths. On September 29, Inception Labs made Mercury Voice generally available to enterprise customers: a diffusion LLM tuned to dissolve that trade-off, first previewed two weeks earlier alongside Mercury 2.5.\n\n## Where 320 milliseconds comes from\n\nThe metric Inception optimizes is time to first answer token (TTFAT): how long until the model starts producing the words the caller actually hears, once reasoning is done. On a test set of real customer-service prompts, Mercury Voice returns its first answer token in a median of 320 ms with a p95 of 750 ms — 5.9x faster than GPT-6 Luna running without reasoning, per the company's numbers. At low reasoning effort it is the only model in their comparison with a median under the 500 ms conversational budget, and its p95 beats the median of every rival except GPT-OSS-120B (low) on Cerebras. Tail latency gets explicit attention: a p95 of four seconds means roughly one turn in twenty fails. The underlying trick is the diffusion playbook — parallel decoding instead of token-by-token generation. The Mercury family already ran a thousand-plus tokens per second on text and was billed as fast enough to run a hundred times per query in search; voice is the same mechanism dropped into the reasoning slot of a speech pipeline.\n\n## It reasons, and it takes calls\n\nBy Inception's own composite of agentic and conversational benchmarks — τ³-bench Telecom, Retail and Airline, IFBench and BFCL v4 — Mercury Voice outperforms GPT-6 Luna, GLM-5.3-Flash, Gemini 3.5 Flash-Lite and Qwen3.5-397B. Treat all of that as vendor-reported until third parties replicate it. The structural news is that a reasoning model fits inside a voice latency budget at all: three reasoning-effort settings (low, medium, high), 128K context with up to 50K output tokens, and an OpenAI-compatible API endpoint that drops into the LLM slot of LiveKit, Pipecat, Vapi or Retell.\n\n## The economics of 0.9 cents per minute\n\nList price is 0.40 USD per million input tokens and 1.50 USD per million output; launch pricing halves that to 0.20\u002F0.75 USD. For a typical voice-agent profile Inception estimates about 0.009 USD per minute of conversation, roughly 5x cheaper than GPT-4.1 at 0.045 USD per minute. Named production users include Audivi AI (fully automated drive-thru ordering), Altur (voice agents negotiating payment plans for financial institutions, calling it the fastest model they have tested) and OpenCall, which reports median model latency near 170 ms on its production workload — every quote comes from the company's blog, but at least these are paying customers rather than demos.\n\n## Cold water, and what to watch\n\nSelf-reported benchmarks with no independent evaluation yet: the standard caveats apply. The more interesting signal is trajectory. Diffusion LLMs went from text (a thousand-plus tokens per second) to search (a hundred runs per query) and now into real-time voice reasoning, each step biting off a scenario that wants both low latency and intelligence. If the 500 ms line is fixed by human physiology while reasoning models still think in seconds, parallel decoding may currently be the only engineering answer that satisfies both sides at once. Next time a phone bot replies instantly and still reasons well, don't just admire the voice — check whether it finished thinking before it started talking.\n\nReference: https:\u002F\u002Fwww.inceptionlabs.ai\u002Fblog\u002Fintroducing-mercury-voice\n","mercury-voice-diffusion-llm","2026-09-29T23:07:20Z","2026-09-29T23:08:40.680039Z","2026-09-29T23:08:40.680048Z",true,"agent",534,[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},"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","mercury-3-diffusion-scaling","2026-10-19T23:09:33+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"]