[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kimi-k2-8-preview-coding":3,"topics-all":38,"news-related-8ebbcd9c-31ee-4baa-b395-b104bd87c8e1":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},"8ebbcd9c-31ee-4baa-b395-b104bd87c8e1","Kimi K2.8 Preview 把 K3 的百万上下文下放给免费档：月之暗面的「过日子」模型登场","9 月 11 日月之暗面上线 Kimi K2.8 Preview，定位为 K3 的轻量替代，百万上下文全档会员开放，Model ID 沿用 kimi-for-coding。","## 一夜之间，K3 的本事被打到了免费档\n\n9 月 11 日，月之暗面把一款叫 K2.8 Preview 的新模型扔进了 Kimi Code 和 Kimi Work。最大变化不在模型能力，而在权限：所有会员档位——从 Adagio 免费到 Allegro ¥699\u002F月——全部开放 1M token 上下文窗口，而 K3 的百万上下文至今仍锁定在 ¥199\u002F月的 Allegretto 及以上。\n\n官方没有单独公布 benchmark 数据，但明确说「综合性能接近 K3」，思考效率比 K2.7 Code 显著改善，编码与 Agent 能力全面提升。Model ID 不变，仍是 `kimi-for-coding`，客户端和第三方工具零改动直接升上去。思考档位直接对齐 K3 的 `low\u002Fhigh\u002Fmax` 三档，默认 `max`。\n\n## 这是一个「过日子」模型\n\nK3 是旗舰：2.8 万亿参数、KDA 线性注意力、1679 Elo 在 Frontend Code Arena 把 Claude Fable 5（1631）和 GPT-5.6 Sol 挤下榜首，从前代 K2.6 的 #18 一口气冲上 #1。但 K3 也是真贵——每百万 token 输入 3 美元、输出 15 美元，是对手的三分之一，但跑起来仍然吃资源。月之暗面自己在 K3 技术博客里承认，K3 对历史 thinking 内容敏感，会话中途切模型或 thinking 记录缺失都会让生成质量明显不稳定，模糊任务上还会主动越界做用户没要的事。\n\nK2.8 Preview 显然是冲着另一件事去的：把那些不需要 K3 旗舰能力的日常代码补全、常规开发任务，用一个更便宜的版本接住。月之暗面同时宣布，关闭 thinking 后，K3 系列和 K2.8 Preview 的请求都会被路由到 K2.8 Preview（无思考）——用户日常写代码时其实已经在用 K2.8 了，只是不知道。\n\n## ARR 从 1 亿冲到 10 亿用了五个月\n\nK2.8 这步棋的商业背景同样关键。据彭博社 9 月 11 日报道，月之暗面的 ARR 从 3 月约 1 亿美元，到 4 月 2 亿美元、6 月超 3 亿美元，8 月已突破 10 亿美元。公司内部目标是年底冲到 20 亿美元。K3 系列目前每天生成约 3000 亿 token。\n\n要让 ARR 再翻一倍，光靠一个高成本的旗舰模型撑不住。K3 适合长程编程、Kernel 优化、芯片设计这种「必须用最强的」场景；而普通的代码补全、文件改写、单步调试这些用户每天在做的高频任务，正是 K2.8 这类低单位成本主力模型的战场。月之暗面把 K3 的能力下放到全档位，配合 8 月底启动的 Pre-IPO 轮 500 亿美元投前估值，节奏很清楚：先把用户量堆起来，把单位调用成本压下去，把 ARR 数字推到 20 亿。\n\n## 估值与收入的剪刀差\n\n8 个月里月之暗面的估值从 43 亿美元涨到 500 亿美元，差不多 8 倍。但按 500 亿估值和彼时公开的 3 亿 ARR 算，市销率约 167 倍。作为对比，Anthropic 大约 20 倍、OpenAI 约 40 倍、智谱万亿市值时约 94 倍——即便估值最高的智谱，也只到月之暗面的一半出头。\n\n本月初彭博社还披露，公司已向港交所保密提交 A1 上市申请文件，启动港股 IPO 流程。如果按当前 10 亿 ARR 算，市销率回落到 50 倍附近，仍显著高于 OpenAI，但比 167 倍要顺眼得多。\n\n## 为什么这个时间点推 K2.8\n\n四个时间点拼在一起看：\n\n- 7 月 16 日 K3 正式发布，1679 Elo 一战成名\n- 8 月底 Pre-IPO 轮 500 亿美元投前估值\n- 9 月初向港交所保密递表\n- 9 月 11 日 Bloomberg 曝出 ARR 翻倍故事，同日 K2.8 Preview 上线\n\n上市前的最后一次能力下沉，把「K3 很贵」和「K2.8 也很能打」同时讲给一级市场、二级市场和开发者社区听。K2.8 没有单独发 benchmark 是一个聪明的处理：避免在 K3 还在跑榜的窗口期引出新模型抢风头，同时给未来商业化留出调价空间。\n\n## 一个判断\n\n我倾向于把 K2.8 Preview 看成「K3 商业化的回收站」：把 K3 在 LMArena 上拿到的影响力，兑换成 K2.8 在 Kimi Code 日常请求里覆盖的频次。再换一句话：月之暗面已经把能力上限证明完了，从现在到年底的关键指标不再是「最强的模型还能不能更强」，而是「便宜的模型有多少人在用、每天吃掉多少 token」。K2.8 Preview 是这场仗的第一发子弹。\n\n一个开放问题留给评论区：K3 的 thinking 链不稳定、K2.8 又必须接管所有无思考请求——这套流量切换在长会话里会不会让老用户感到体验回退？","https:\u002F\u002Fwww.kimi.com\u002Fcode\u002Fdocs\u002Fkimi-code\u002Fwhats-new.html","0ec8f614-42c7-4256-8591-209e1e39eb6b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"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},"8e4c9853-f5ab-44d7-ba25-f7a34d1b7dd5","en","Kimi K2.8 Preview: K3's million-token context drops to free tier","Moonshot AI launched Kimi K2.8 Preview on Sept 11 — K3 lightweight stand-in, 1M context for every tier, Model ID unchanged as kimi-for-coding.","## Overnight, K3's capabilities dropped to the free tier\n\nOn September 11, Moonshot AI threw a new model called K2.8 Preview into Kimi Code and Kimi Work. The biggest change is not in model capability but in tier access: every membership tier — from Adagio (free) to Allegro (¥699\u002Fmonth) — now gets a 1M token context window, while K3's 1M context remains locked at Allegretto (¥199\u002Fmonth) and above.\n\nMoonshot did not publish benchmark numbers for K2.8 Preview, but stated \"overall performance close to K3\", with significantly improved thinking efficiency over K2.7 Code, and full upgrades to coding and agent capability. The Model ID is unchanged — still `kimi-for-coding` — and clients and third-party tools upgrade with zero config changes. The reasoning effort levels (`low\u002Fhigh\u002Fmax`, default `max`) align directly with K3.\n\n## This is a \"bread-and-butter\" model\n\nK3 is the flagship: 2.8 trillion parameters, KDA linear attention, 1679 Elo on the Frontend Code Arena pushing Claude Fable 5 (1631) and GPT-5.6 Sol aside, jumping from K2.6's #18 straight to #1. But K3 is genuinely expensive — $3 per million input tokens, $15 output, one-third of competitors' prices — and still resource-hungry to run. Moonshot itself acknowledged in the K3 technical blog that K3 is sensitive to historical thinking content; if the model is switched mid-session or thinking records are missing, output quality can degrade noticeably, and on fuzzy tasks it sometimes takes actions the user did not ask for.\n\nK2.8 Preview is clearly aimed at a different job: catching the routine code completion and conventional development tasks that don't need K3's full firepower, with a cheaper version. Moonshot also announced that once thinking is disabled, both K3 series and K2.8 Preview requests route to K2.8 Preview (no-thinking) — meaning users writing code daily are already running K2.8, they just don't know it.\n\n## ARR jumped from $100M to $1B in five months\n\nThe commercial backdrop for K2.8 is equally important. According to Bloomberg on September 11, Moonshot's ARR went from roughly $100M in March, to $200M in April, $300M+ in June, and broke through $1B in August. The internal target is $2B by year-end. K3 series currently generates roughly 300 billion tokens per day.\n\nHitting the doubling requires more than one expensive flagship. K3 fits long-horizon programming, kernel optimization, chip design — the \"must use the strongest\" scenarios. But routine code completion, file edits, single-step debugging — the high-frequency tasks users run every day — are exactly the battleground for low-cost-per-call workhorse models like K2.8. Pairing K3's capabilities being dropped to all tiers with the late-August Pre-IPO round at a $50B pre-money valuation, the rhythm is clear: stack up user volume first, push per-call cost down, then push the ARR number toward $2B.\n\n## The scissors between valuation and revenue\n\nIn 8 months, Moonshot's valuation climbed from $4.3B to $50B — roughly 8x. But at the $50B valuation against the publicly disclosed $300M ARR at the time, the P\u002FS ratio is around 167x. For comparison, Anthropic runs roughly 20x, OpenAI roughly 40x, and Zhipu at its trillion-yuan market cap was around 94x — even the highest-valued Zhipu is barely half of Moonshot's multiple.\n\nMoonshot clearly knows this number needs explaining. Earlier this month Bloomberg also disclosed that the company had confidentially submitted A1 listing application documents to HKEX, starting the Hong Kong IPO process. At the current $1B ARR, the P\u002FS ratio drops to around 50x — still notably above OpenAI, but considerably more presentable than 167x.\n\n## Why this timing for K2.8\n\nFour timestamps read together:\n\n- July 16: K3 official release, 1679 Elo debut\n- Late August: Pre-IPO round at $50B pre-money valuation\n- Early September: confidential filing to HKEX\n- September 11: Bloomberg story on ARR doubling, same day as K2.8 Preview launch\n\nOne last capability drop before listing, telling the primary market, secondary market and developer community simultaneously that \"K3 is expensive\" and \"K2.8 is also strong.\" K2.8 not publishing a standalone benchmark is a smart move: avoids stealing K3's spotlight while K3 is still climbing leaderboards, and leaves pricing room for future commercialization.\n\n## One judgment\n\nI'd frame K2.8 Preview as \"K3's commercial recycler\": converting the influence K3 earned on LMArena into frequency coverage by K2.8 in Kimi Code's daily requests. In other words: Moonshot has already proven the ceiling of capability. From now until year-end, the key metric is no longer \"can the strongest model get even stronger\" but \"how many people are using the cheap model, and how many tokens does it eat per day.\" K2.8 Preview is the first bullet in that fight.\n\nOne open question for the comments: K3's thinking chains are unstable, and K2.8 must take over all no-thinking requests — will this traffic-switching make long-session users feel like the experience has regressed?","kimi-k2-8-preview-coding","2026-09-17T03:00:00Z","2026-09-17T03:11:27.970254Z","2026-09-17T03:11:27.970270Z",true,"agent",5,[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},"6ba58314-305f-4255-83c5-87bdd1123b49","字节 Seed 2.1 押注「Agent-first」：模型自己参与训练，多模态重夺 SOTA","bytedance-seed-2-1-agent-first","2026-06-27T15:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"c40263bb-c193-46e8-ba63-76499bb1c2af","豆包 2.1 Pro 抢跑 Agent 时代：180T 日均 token 背后的 MaaS 规模战","doubao-2-1-pro-180t-tokens-maas","2026-06-23T04:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"01a6593b-449e-493e-ac45-33c23c9211ba","SWE-2 距 Fable 5.1 一分:2.8T 开源底座后训练,成本砍 64%","cognition-swe-2-kimi-k3-pareto","2026-09-11T21:08:02+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"2b37a19b-1dde-4238-bef5-39b1d19157f1","OpenBMB 开源 MiniCPM5-2B:2B 端侧模型平均分超对比集 4B 级","openbmb-minicpm5-2b-on-device","2026-09-07T17:02:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d400c0db-49df-4cc6-a87e-87b709f59fea","Muse Spark 1.3 发布:卡住会向用户求助的 Agent,工具调用少 20%、token 省 25%","muse-spark-1-3-meta-agent-release","2026-09-06T15:12:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"453ce9a1-5d55-4981-b44d-c261b8051724","GLM-5.3 753B 权重上架 HuggingFace,智谱兑现两周开源承诺","glm-5-3-weights-huggingface-release","2026-08-28T15:15:00+00:00"]