“全球大模型第一股”来了!“中国OpenAI”智谱正式在港交所挂牌上市 - 中国科技网
中国何时出最强大模型?智谱回应马斯克:比你想得快 - 观察者网
What if diffusion models generates Neural Network weights? Gen AI generate Gen AI [R]
| I'have been dreaming about using a Diffusion Model that will generates Weights instead of pixels. and in the future more complex scenarios! I made, with the help of z.ai zcode and their fantastic GLM model a small POC that i published on my github: It's a very small experimental implementation , but somenthing seems to work! I have not find paper on similar implementation ad I'm asking the community to help me to find one, and to have contribution on this . I also made a small webapp in order to test the trained model (actually i'm only on half training path 3000/4000 steps finished, but i'm using my poor laptop CPU at the moment) This is some screenshot of preliminary results. (the best results achieved, in truth) Italian explanation on my YouTube https://youtu.be/wAMO3Db-JoQ [link] [comments] |
SCDBench: A Benchmark for LLM-Based Smart Contract Decompilers
I Think, Therefore I Attack: Evaluating Thought Injection Attacks on Reasoning Agents
Current Reward Hacking Model Organisms Don’t Hack Like Real Models
Summary
- Current reward hacking model organisms are poor proxies for real models. They do not exploit the hacks that open-weight frontier models do, and mostly hack narrowly or not at all.
- This is based on our evaluation of three model organisms: Mini HackGPT (GPT-OSS-120B from Golechha et al., 2026), HackGPT, and HacKimi (GPT-OSS-120B and Kimi K2.5 from Yudelson et al., 2026 respectively). We ran these organisms in MiMo V2.6 RL environments we modified by inserting hacks such as git history leaks, a tamperable grader, and a forbidden directory containing tests.
- In these environments, HackGPT only ever hacks by tampering with the grader, while Mini HackGPT and HacKimi never hack. In contrast, current open-weight models like DeepSeek V4.1 Flash 1017, GLM 5.3, and MiMo V2.6 Pro never tamper with the grader, but are able to exploit the other hacks we inserted.
- This suggests that open-weight models are in fact better reward hackers than existing model organisms. This is problematic because it means we do not have usable model organisms to help us understand reward hacking behavior and test interventions to mitigate it — existing model organisms are not representative of frontier models, while the frontier models themselves do not have pre- or intermediate RL checkpoints we can use to run controlled experiments.
Introduction
Reward hacking encompasses many different types of related behavior, and it is currently poorly understood how these are linked to each other. Despite the best efforts of frontier labs to align, audit, and evaluate models, we still routinely see many misalignment incidents occurring after they are deployed in the real world, which have been attributed to models learning to reward hack during RL.
In this work, we evaluate what we consider to be the three most useful reward hacking model organisms: Mini HackGPT (GPT-OSS-120B from Golechha et al., 2026), HackGPT, and HacKimi (GPT-OSS-120B and Kimi K2.5 from Yudelson et al., 2026 respectively). These model organisms were trained on RL environments primarily containing hacks that could be exploited by modifying the grader or tests. We test whether their reward hacking behavior generalizes by running them on environments with inserted hacks that require more than just test or grader gaming to exploit, such as git history leaks and hidden tests. We think this is important as it informs how effectively we can use the model organisms to understand what causes reward hacking and test alignment techniques intended to mitigate it.
Setup
Model organisms
We tested three existing reward hacking model organisms from Golechha et al. (2026) and Yudelson et al. (2026). Note that their training environments consist mostly of hacks that would encourage test gaming. These are also currently the largest and most recent model organisms that have been trained for studying reward hacking.
Organism | Training environments | Hacks in training environments | Source |
HackGPT | Various coding environments detailed here: Our Reward Hacking Envs ~50% of environments are a fully Claude-generated, not ground-truth-tested set of coding problems with test suites |
Zhipu.AI’s Open-Source Power Play: Blazing-Fast GLM Models & Global Expansion Ahead of Potential IPO
Beijing, China – April 15, 2025 – In a strategic move that underscores its technological prowess and global ambitions, potentially paving the way for a future IPO, Chinese AI company Zhipu.AI has announced the comprehensive open-sourcing of its next-generation General Language Models (GLM). This release includes the advanced GLM-4 series and the groundbreaking GLM-Z1 inference models, boasting unprecedented inference speeds and the launch of a dedicated international domain, Z.ai.
The spotlight shines on the GLM-Z1 inference model, which Zhipu claims achieves inference speeds up to eight times faster than DeepSeek-R1. By optimizing GQA parameters, employing quantization, and implementing speculative sampling, the GLM-Z1-32B-0414 delivers a remarkable 200 tokens per second on consumer-grade GPUs – a staggering 50 times faster than human reading speed. This exceptional responsiveness positions it as a frontrunner in efficient AI inference, a key advantage as Zhipu eyes further market expansion.
Further demonstrating its innovation, Zhipu has unveiled the “Rumination” model, GLM-Z1-Rumination-32B-0414. This model signifies a leap towards more autonomous AI agents, capable of actively searching the internet, utilizing tools, conducting in-depth analysis, and self-verifying information to tackle complex, open-ended queries – a significant step beyond purely reactive AI and a testament to Zhipu’s cutting-edge research.
The open-sourced portfolio also features the foundational GLM-4-32B-0414, specifically enhanced for agent capabilities with superior performance in tool usage, web search, and code generation. Its real-time code generation capability for languages like HTML, CSS, JS, and SVG directly within conversations offers a significant boost to developer productivity.
Recognizing the diverse needs of the AI community, Zhipu has also open-sourced smaller 9B parameter versions of both GLM-4 and GLM-Z1 models. These compact yet powerful models demonstrate impressive performance in mathematical reasoning and general tasks, providing an efficient solution for resource-constrained environments and further broadening Zhipu’s appeal. All models are released under the permissive MIT license.
This strategic open-sourcing initiative, coupled with the launch of the international-facing Z.ai platform, strongly signals Zhipu.AI’s commitment to global accessibility and fostering a vibrant open-source AI ecosystem. The new domain, Z.ai, serves as a central hub for users worldwide to freely experience these advanced models through a web interface and dedicated app.
For enterprise clients, Zhipu continues to offer its robust Model-as-a-Service (MaaS) platform, now integrating the newly open-sourced base and inference models. This platform provides API access with tiered pricing, including the ultra-fast GLM-Z1-AirX, the highly cost-effective GLM-Z1-Air, and the free GLM-Z1-Flash, catering to a wide range of commercial applications. The foundational models GLM-4-Air-250414 and the free GLM-4-Flash-250414 are also available on the MaaS platform.
As Zhipu.AI strategically expands its global footprint and showcases its technological leadership through this significant open-source release and the launch of Z.ai, the move could be interpreted as a strong indicator of the company’s readiness and ambition for a potential IPO in the near future. By democratizing access to its cutting-edge AI technology, Zhipu is not only fostering innovation but also building a strong global presence and user base.
Explore the new GLM models for free at: https://chat.z.ai/ Open Source Download: https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e
The post Zhipu.AI’s Open-Source Power Play: Blazing-Fast GLM Models & Global Expansion Ahead of Potential IPO first appeared on Synced.
刘庆峰交出科大讯飞半张AI答卷
(本文作者为 唐辰同学,钛媒体经授权发布)
文 | 唐辰同学
科大讯飞的AI预期兑现了多少?
4月29日的业绩说明会上,科大讯飞董事长刘庆峰主动提起了市值。他说,对很多投资人来说其实没达到期望,“我们也觉得有点抱歉”,“公司的市值应该做得更好,我们要把国产芯片这条艰难的路让大家看透,在商业逻辑上让大家看清楚。”
他的潜台词是,走全国产算力路线的讯飞,市值被低估太多,对应当天市值约为1154亿元。
也是在那场会上,刘庆峰给国产算力的困境定了期限。他说,训练超长上下文确实受限,但困难 “已经看得很清晰”。讯飞正与华为相关团队深度配合,预计10月出成果,阵痛就会过去。
讯飞兑现技术预期
半年过去,他开始交作业了。
10月22日下午,1024开发者节首次在北京举行。刘庆峰发布端云一体的星火X2.6,首次在国产芯片上跑通百万级超长上下文的代码与智能体训练,超长上下文训练效率提升约4倍。
三天前的10月19日盘后,讯飞2026年三季度财报出来了。期内公司实现营业收入65.64亿元,同比增长7.99%;归母净利润2.26亿元,同比增长31.05%;前三季度累计归母净利润2188.72万元,同比扭亏。
不过扣除非经常性损益后,第三季度依然亏损1.26亿元,前三季度累计亏损7.63亿元。
两个季度之后,这条路的商业逻辑还在被检验,其市值变为878亿元,比年初的阶段高点跌去四成多。
技术推进顺利,财务表现差点意思。这个落差,就是外界对科大讯飞的AI预期:一半是技术,一半是商业。
技术这一块,看国产算力能不能撑起主流大模型,代码和智能体跟不跟得上;商业那一半,看技术落地后钱能不能进来,市值认不认。
刘庆峰很好地兑现了前一半;后一半,尤其是市值,还悬着。
换句话说,他交出来的,是科大讯飞AI的半张答卷。
大模型这三年,刘庆峰在技术上没画大饼,设定的目标基本都完成了。比如,2023年5月6日,星火大模型1.0发布,他放话10月份对标ChatGPT(GPT 3.5),还给出了年内三次升级的时间表:补数学、补代码,中文要超、英文要平。
五个月后,星火V3.0发布,刘庆峰说:“我们如期兑现了自己给大家的时刻表。”
2026年的主角是Agent,这轮竞争最吃算力的是代码和智能体。讯飞的解法是硬攻:训练端把长任务拆开并行跑,按任务长短分配算力;推理端靠多层缓存和分布式,破局“装不下、训不动”。
他在主论坛透露,端侧模型能独立完成的智能体任务,覆盖度从行业平均的12.8%提到41.9%。训练效率上,长思维链强化学习的效率从30%提到84%,MoE全链路从30%提到93%。
其代价是每个国际主流算法出来,讯飞都没法直接用,得先花三到六个月在国产卡上把它适配、跑通。这个时间成本最后都落在研发费用上,一年比一年重。
相较A股软件公司研发费用率普遍在15%–20%,科大讯飞连续多个季度的研发费用率在25%左右。
但尴尬的是,技术上的“交付”,对公司营收、市值的拉动作用,并没有达到预期。
2021年2月,刘庆峰在年度计划大会上,给人工智能2.0时代定下目标:“十亿用户、千亿收入、万亿生态”。完成时间定在“十四五”收官的2025年,千亿收入被拆成七个板块的指标,加起来正好1000亿元。当时讯飞一年营收只有130.25亿元,等于五年内营收要翻七倍半。
2025年,目标到期。讯飞完成营收271.05亿元,比五年前翻了一倍。这个增速放在A股,不算慢,只是对照千亿收入目标,差距就过于明显。
收入之外的利润构成也很有意思。2025年归母净利润8.39亿元,扣掉非经常性损益,只剩2.64亿元:三成出头是主业挣的,其余的大头,是计入损益的6.36亿元政府补助。
这笔补助对应它承担的国家项目,不是白拿的,但确实垫高了账面利润。
利润表之外,还有应收这笔钱。2025年末,讯飞账上162.96亿元应收账款,相当于全年营收的六成,其中56亿元账龄超过两年。
折算下来,讯飞每做10块钱生意,6块还没收回来,其中有2块已经在账本挂了两年以上。像教育、政务这类生意,交付和付款周期本来就长,项目越多垫得越多。
讯飞是赚钱的,只是回款周期比较长。
讯飞有没有被低估?
刘庆峰挖的护城河是国产算力,这既是他的护身符,也是他的天花板。
2023年1024开发者节,讯飞与华为联合发布国内首个全国产万卡平台“飞星一号”。刘庆峰反复强调:别的公司只在推理阶段适配国产芯片,讯飞从训练端就开始用。
这是其卖点,但短板也客观存在。今年8月的业绩说明会上,讯飞承认在256K以上超长上下文、复杂智能体和代码训练上,与英伟达H200有约五倍差距。
刘庆峰说,“因为国产算力对超长上下文的限制,使得在这一波以代码和智能体为主的市场热点中,讯飞看起来是有所缺位的。”
他肯把差距摆在桌面上讲,很不容易。全国产这块招牌,国内没有第二家。其底气还是来自华为的新卡,昇腾950DT对标H200,综合性能已经做到65%到85%,公司预计今年底至明年显著改善。
不过,这个差距在刘庆峰看来伤不到基本盘:教育、医疗、汽车、央国企这些行业应用里,90%甚至95%以上的场景用不到那么长的上下文,几十万Token就够了。
如果接下来市场比的不是参数,而是Token,讯飞这笔算力投入还算不算划算?
刘庆峰表示,“真正在API和纯Token消耗获得的收益,我们已经比其他模型公司高很多了。”
这在财报里得到验证,讯飞披露,2025年,其大模型API及MaaS收入3.85亿元,同比增长263%。
技术交了完整答卷,市场为什么不认?这不是讯飞一家的事。
年初上市的智谱、MiniMax也都从高点回落过,变化在定价方式上:以前听故事,现在算Token、看ARR(年度经常性收入)。最典型的指标叫P/ARR,也就是市值除以年度经常性收入。它算的是市场愿意为一块钱“能持续收进来的钱”出多少价。
这套估值算法适合API按调用量持续计费的纯模型公司。以智谱为例,它2025年总收入7.24亿元。但市场给它定价,参考是18亿美元的ARR。
一个是全年确认收入,一个是年度经常性收入。讯飞就算想把这个差别抹平,量级还是差着几倍——讯飞一整年的API及MaaS收入,大约只等于智谱十来天的ARR。
而且这还是拿讯飞最像模型公司的一块业务,去比人家的全部收入。
所以,在市场现在最认可的Token收入上,刘庆峰说的“高很多”,是增速上的领先,不是量级上的。
在讯飞开放平台,日均Tokens一年涨了42倍,收入暂时还只有几个亿。
显然,Token这门生意不是讯飞的主场,它的大头在教育、医疗、政企这些行业。其商业模式是项目制,需要讯飞一个个派团队、一家家攒数据、一年年做交付,是软件公司的路子。
2025年大模型相关项目中标金额23.16亿元,数量和金额都是行业第一,超过第二到第六名的总和。
或者说,市场对讯飞的“成见”是,把两种生意当成一种来定价:收入按Token计费的公司,享受高倍数;收入按项目结算的,被当成传统集成商。
刘庆峰的转变
市场的定价方式短期改不了,能改的只有讯飞自己。8月21日下午的半年度业绩说明会上,刘庆峰做了三个与他过去做法相反的表态。
首先是当众承认讯飞慢了。他自己捅破“有所缺位”的窗户纸,还称“今年上半年我们感到对投资者有很大的遗憾和压力”;科大讯飞董秘江涛也很坦诚,他从营收、利润到现金流,“很多投资者都对这份半年报表达了不满意”。
一家A股上市公司的管理层,把投资者的不满原样摆到台面上讲,这不算常见。
其次是讯飞的算力从买改成租,要么直接租,要么与政府、相关厂商合建之后再租,不再一次性砸大钱。
第三,他给烧钱划了条线:星火大模型已经拆出去独立成公司,接下来引入外部融资;即便拿到外部资金,也不会不计成本投入算力,去做单纯的模型“摸高”。
正视现实、改换模式、踩刹车,这与他过去摸高创收、自己买算力、猛砸钱搞研发的风格迥异。
讯飞和刘庆峰都认识到,技术兑现之后,商业预期也要开始兑现了。
从三季报和半年报的账面数据可以看到,讯飞拿到了不少“好结果”。比如收入结构在切换,上半年开放平台收入37.05亿元,同比增长36.01%,第一次超过智慧教育的34.90亿元。这说明,讯飞的AI新业务正在起势,在接过智能硬件、智慧城市这些老业务的增长重任。
研发投入也还在增加。上半年研发费用30.07亿元,同比多花6亿多,对应的扣非亏损是6.37亿元。
这释放的一个信号是,讯飞的亏损是因为投入在不断增加,不是其自身的商业模式有问题。
在此之前,刘庆峰就表达过相关态度,他说,加大研发投入是出于对大模型未来的信心。“如果我们选择保守,把这些投入都省掉,好像利润马上就出来了”。

一进一出,缺口只能靠融资补:前三季度经营现金流净额2.64亿元,同比增长115.28%,回款确实在变好;但同期资本开支34.78亿元,自由现金流仍是负32亿元,靠筹资净流入60.59亿元顶上,其中定增约40亿元。
定增能顶一时的缺口,但要真正让市场在Token、ARR、扣非净利润、自由现金流四个指标上,对它重估,讯飞眼下还有很长的路要走。
刘庆峰还是相信自己这些年攒下来的家底:国内独一份的全国产算力训练出来的主流大模型;20多年来在教育、医疗、政法里沉淀的数据和行业理解;学习机、翻译机、办公本、AI眼镜、汽车座舱这一串软硬一体的终端。
而且,他认定,“绝不是所谓一人公司拿着工具就能做出来”的。
我注意到,外界看他这堆家底,有个说法:讯飞AI布局像个杂货铺。但我不认为这是差评,因为基于国产算力,把这些东西一件件做出来,正是头部AI公司应该具备的开拓能力。
但技术兑现与商业预期的错位,还是给刘庆峰提出了一个绕不开的问题:讯飞什么时候能靠主业赚钱?
他在4月份说过:“我相信,未来我们的路越走越宽时,我们自己一定会笑在最后,投资人也会笑到最后。”
在我看来,讯飞和投资人要笑到最后,还是得明确一下自己的主业,以及对自身体系的重新审视。刘庆峰也才能完成科大讯飞AI的完整答卷,并继续兑现他定下的调子:“跟所有的顶尖玩家站在同一个平台上”。
参考资料:
XAI财经,《科大讯飞的AI预期,兑现了多少?》
钛媒体,《科大讯飞刘庆峰:讯飞星火X2.6全国产算力训练》
更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App
消费三姐妹,为什么不香了?
(本文作者为 定焦One,钛媒体经授权发布)
文 | 定焦One(dingjiaoone),作者 | 陈丹,编辑 | 魏佳
一只玩偶、一件金饰、一杯奶茶,曾在港股撑起超过7300亿港元的市值。
2025年6月,泡泡玛特、老铺黄金、蜜雪集团股价先后创下新高,三家公司市值合计一度超过7300亿港元,成为资本市场炙手可热的“消费三姐妹”。一年多过去,三家公司股价均较各自高点跌去一半以上,其中老铺黄金跌幅接近四分之三。
最新的坏消息来自泡泡玛特。10月27日,公司披露第三季度经营数据,整体收益同比下降45%至50%,美洲市场降幅高达70%至75%。而一年前,LABUBU正席卷全球,同期收益还在以245%至250%的速度增长。增长急转直下,次日开盘,泡泡玛特股价一度跌超6%。
另外两家,即使利润仍在大幅增长,也没能阻止股价下跌。
老铺黄金今年上半年收入增长60.3%,经调整净利润增长83.6%,股价却是三家公司中回撤最深的。蜜雪集团同期收入增长2.3%,净利润下降14.7%,同样未能摆脱股价下行的趋势。
潮玩、金饰、平价茶饮,三门生意的增长曲线并不相同,却遭遇了相似处境。去年,资本市场愿意为爆款IP、高端品牌和连锁扩张带来的增长想象支付高价;如今,即便仍有增长,投资者也不再愿意给予同样高的估值。
与此同时,AI算力、半导体等科技板块成为市场新的投资热点。与增长逐渐回归常态的消费公司相比,AI产业链的需求扩张和技术突破,为投资者提供了另一种关于未来增长的想象。
消费三姐妹的集体失宠,究竟是自身业绩难以支撑过去的高估值,还是市场找到了更有吸引力的投资方向?
01.从消费到AI,资本市场换了新宠
2025年,消费三姐妹的走红,有着实打实的需求和业绩支撑。香颂资本董事沈萌认为,这轮上涨受到国内消费需求结构性变化的驱动。悦己、保值、低价等消费偏好形成新的趋势,带动相关企业快速增长。三家公司的崛起,恰好反映了这些不同的消费需求。
2025年上半年,国际金价屡创新高,老铺黄金门店也掀起排队潮,上海豫园门店一度排起长达10小时的队伍。消费者愿意为古法工艺、设计和品牌支付溢价,也看重足金材质的价值支撑。同年,LABUBU从柜子里的收藏品变成年轻人挂在包上的时尚配饰,热潮一路蔓延至海外。一个卖黄金,一个卖玩偶,看似毫不相关,却都让消费者愿意为产品本身之外的价值付费。
另一边,蜜雪冰城靠几块钱的产品做大生意。2元的冰淇淋、4元的柠檬水,背后是大规模采购、自建生产和物流配送形成的成本优势。到2025年6月底,蜜雪集团全球门店达到53014家,其中内地超过一半的门店位于三线及以下城市。
这三种看似不同的消费选择,都带来了可观的业绩增长。2025年,泡泡玛特毛绒产品收入同比增长560.6%至187.08亿元,海外收入增长291.9%;老铺黄金上半年收入增长251%,净利润增长285.8%;蜜雪集团同期收入增长39.3%。
三家公司分别找到了愿意为情绪价值、品牌与黄金材质、低价商品买单的消费者,这种增长的稀缺性,正是当时资本市场愿意支付高估值的重要原因。
2025年8月,泡泡玛特股价一度逼近340港元,市值超过4500亿港元,市盈率超过105倍。老铺黄金的市盈率一度达到约140倍。即便是三家中市盈率相对低的蜜雪集团,也超过45倍。100倍市盈率,意味着每1元年度净利润对应100元市值。投资者愿意出这样的价格,买的显然不只是眼前的利润,还有未来多年持续增长的预期。
在沈萌看来,资本市场往往会为短期股价上涨寻找合理解释,将企业的阶段性优势包装成具有长期成长性的投资故事。但被赋予某种概念,并不意味着企业真正具备持续兑现这一预期的能力。
风险很早就开始显现。2025年6月5日,港股新消费龙头集体回调,老铺黄金单日下跌9.05%,蜜雪集团下跌7.72%。就在前一天,三家公司还曾齐创股价新高。当时,富国基金经理周文波就提醒,新消费公司吸引了大量资金关注,但部分公司已不便宜,短期股价存在调整风险。
此后,三家公司的股价走势有所分化,泡泡玛特还在8月创下更高纪录。但市场对高估值能否持续的担忧,已经持续出现。而从2025年下半年开始,AI成为资本市场更受关注的增长方向。随着资金向算力、半导体等科技板块集中,前期热门消费股在公募重仓中的热度逐渐降温。
据Choice统计,重仓泡泡玛特的公募基金数量从2025年二季度末的311只一路缩减,2025年四季度末降至136只,2026年一季度末为110只,到2026年二季度末只剩86只,累计减少逾七成;持股数量从2026年一季度末的3324.82万股降至二季度末的2114.68万股,一个季度减少约36%。
与此同时,AI产业链逐渐成为公募基金的配置重点。据国信证券统计,2026年二季度基金重仓股中,科技板块整体持股市值占比上升23.6个百分点至60.8%;作为科技板块内部占比最大的行业,电子行业持股市值占比上升21.2个百分点至42.9%。消费板块占比则下降7.6个百分点至9.5%。
这种变化,在同一只基金的调仓中表现得更加直观。2025年二季度,中加新兴消费基金的前十大重仓股中,还有泡泡玛特、匠心家居、古茗等消费公司。到了2026年一季度,寒武纪、兆易创新等科技公司已经进入其前十大重仓股。基金经理在季报中解释,科技改变生活仍是全年的产业趋势主线。
申万宏源在2026年二季度研报中指出,科技赛道对其他行业存在资金虹吸效应。一方面,科技硬件吸引了其他行业的配置资金;另一方面,科技主题基金凭借业绩表现,吸引了更多投资者申购。

沈萌认为,这与市场流动性和资金轮动有关。“市场流动性的规模有限,资金在不同热点板块之间轮动。资金都在追逐尽可能安全、尽可能高的回报,而最关键的风向标就是政策导向。”
但资金流向AI,并不意味着AI制造了消费股的下跌。
投资人陈默默认为,从时间线来看,消费三姐妹的高估值在AI资金虹吸效应凸显之前,就已经出现调整压力。AI行情加剧了资金分流,可能让消费股跌得更深、更快,却不是下跌的根本原因。
“钱搬家放大了下跌,没有制造下跌。”陈默默说,即使没有AI,资本市场也会寻找其他具有更高增长预期的资产,消费股仍可能面临调整,只是幅度可能更小、速度也更慢。
到了2026年三季度,公募基金对电子、通信行业的配置也有所回落,医药和部分消费行业重新获得增配。资金的偏好仍在变化,但对消费三姐妹而言,想重新获得市场青睐,要回答的是自己的增长还能走多远。
02.三种生意,撑不起原来的估值了?
同一个问题,三家公司面对的是三种不同的考验。
泡泡玛特的难题是,一个超级IP究竟能火多久。
2025年,THE MONSTERS系列收入141.6亿元,同比增长365.7%,占泡泡玛特总收入的近四成。风靡全球的LABUBU,就出自这一系列。
如今,这股热度正在消退。2025年6月,随着官方集中补货,LABUBU部分款式的二手价格迅速回落。2026年6月,LABUBU 4.0“复古理发店”系列上市,市场反应已明显不如从前。德意志银行的渠道调研显示,该系列电商首日销售表现仅为2026年此前主要IP新品首发时的30%至50%,部分常规款的二手价格甚至跌破159元的官方售价。
新IP仍在增长,却尚未填补整体增长放缓留下的缺口。2026年上半年,星星人IP收入同比增长580.6%。但在8月的业绩会上,创始人王宁坦言,公司今年大概率无法完成年初设定的20%增长目标。
陈默默认为,市场当初并不是押注LABUBU能成为一棵“常青树”,而是期待泡泡玛特成为一个持续生产爆款的IP孵化器,不断推出新的IP,再借助全球化渠道进行分发。但到目前为止,这种持续复制爆款并借助势能持续扩大分发的能力,还有待进一步验证。
老铺黄金的难题在于利润能否持续、稳定地增长。
2026年上半年,老铺黄金收入198.08亿元,归母净利润42.67亿元,同比增长88.2%。但拆分到季度,业绩波动就显现出来。今年2月底,老铺黄金将产品价格上调约20%至30%,涨价前引发一轮集中购买。根据一季度业绩预告区间和半年报倒算,二季度净利润约为4.67亿至6.67亿元,远低于一季度的36亿至38亿元。
董事长徐高明在业绩会上解释,二季度国际金价大幅回调,叠加消费市场转弱,公司选择暂时观望,没有急于调整经营策略。他表示,二季度月均利润仍超过2亿元,符合公司预期。
券商的判断则更为谨慎。2026年8月,开源证券将老铺黄金2026年至2028年的归母净利润预测,分别从90.12亿、123.05亿、155.37亿元,下调至65.35亿、67.43亿、80.03亿元。按照调整后的预测,老铺黄金2027年的利润增幅仅约3%。
老铺黄金的品牌溢价仍在,海外收入保持快速增长,金器等新品类也在贡献增量。但过去的高速增长,部分受益于金价上涨和调价前的集中购买。接下来,它需要证明,即使没有这些因素的推动,复购、新品和渠道扩张仍能支撑业绩增长。
蜜雪集团的难题则是,门店越开越多,利润不增反降。
2026年上半年,蜜雪集团净利润23.19亿元,同比下降14.7%,首次在上市后出现半年利润同比下滑。截至6月底,集团全球门店达到63987家,同比增长约20.7%,但同期收入仅增长2.3%。
开店速度也在放缓。上半年,蜜雪集团新开加盟店5455家,比去年同期减少2266家;净增加盟店4166家,低于去年同期的6534家。
管理层在业绩沟通会上承认,上半年集团整体及蜜雪冰城主品牌的店均营业额均出现两位数下滑。按商品及设备销售收入和平均加盟店数量测算,单店半年向总部的采购额同比下降约17.9%。
这一变化触及了蜜雪集团商业模式的核心。集团收入主要来自向加盟店销售原材料、包材和设备。过去,门店越多,总部卖出的货就越多;但随着门店不断加密,单店生意更容易被分流,新增门店带来的收入增量也可能逐渐缩小。
与此同时,费用仍在上升。上半年,蜜雪集团销售及分销开支增长22.9%,行政开支增长39.4%,均明显快于收入增速。公司解释,部分支出用于产品品质升级、品牌建设和门店运营支持。
对蜜雪集团来说,问题已经转向“每家店还能赚多少钱”。总部收入能否持续增长,最终取决于加盟商能否保持盈利,以及是否愿意继续开店。
沈萌总结说,成本优势容易受到上游价格波动影响,而情绪价值带来的差异化优势,则需要企业不断推出新的产品和体验。它们的竞争优势并未消失,泡泡玛特仍有IP矩阵,老铺黄金仍有品牌溢价,蜜雪集团仍拥有六万多家门店和成熟的供应链体系。改变的不是这些生意本身,而是资本市场对其未来增长的预期。
“虽然业绩表现不同,但投资者对它们未来的预期基本一致,即不足以支撑起当前的估值。”沈萌说。
03.AI接过了溢价,也接过了同一道考题
那么,成为资本市场新宠的AI资产,估值是否更有底气?
从目前来看,答案并不一致。在AI产业链上,有的公司已经将需求转化为利润,有的仍处于高投入阶段,资本市场对它们的定价也出现明显分化。
先看大模型公司。
2026年1月,智谱和MiniMax先后登陆港股。今年上半年,智谱收入9.54亿元,同比增长399.7%,经调整净亏损19.64亿元,亏损同比扩大;MiniMax收入约1.17亿美元,同比增长283%,经调整净亏损2.93亿美元,同比扩大111.2%。
两家公司的收入都增长了数倍,经调整净亏损却仍是同期收入的两倍以上。模型训练、推理和研发需要持续投入,快速增长的收入尚未转化为盈利。资本市场的态度也发生了变化。两家公司上市后,股价一度大幅上涨;进入2026年,又相继经历深度回撤。截至10月20日,智谱和MiniMax股价均较各自历史高点下跌超过八成。
相比之下,资金更青睐已经兑现利润的AI硬件企业。
华夏基金基金经理李湘杰此前指出,当前投资重点集中在上游硬件,下游应用仍处于分化观察阶段。沈萌也认为,大模型故事的热度正在消退,短期内资本市场更关注业绩已经兑现的AI硬件企业。
中际旭创就是代表。2026年上半年,公司收入417.78亿元,同比增长182.49%;归母净利润136.51亿元,同比增长241.70%。AI大客户持续扩建数据中心,带动800G、1.6T高速光模块出货增长。
需求端也仍在扩张。Gartner今年9月预测,2026年全球AI支出将达到2.67万亿美元,同比增长49.5%,其中AI基础设施是最大的支出方向。
部分AI硬件企业正处于收入和利润双双高速增长的阶段,但不代表估值可以无限上涨,一旦增长放缓,一样会被重新计算,这正是消费三姐妹在经历的事。
机构的调仓已经释放出一些信号。
据华福证券统计,2026年三季度主动权益基金对电子、通信行业的重仓配置比例继续回落,环比分别下降6.81个和1.69个百分点,电子配置比例降至约36%,通信降至约15%。
中际旭创仍是公募基金第一大重仓股,但持有它的基金数量和持股数量,均较二季度末有所减少。电子、通信仍是公募基金的重要配置方向,机构并未全面撤离AI。只是随着股价上涨、持仓日益集中,部分资金已经开始减仓,重新权衡未来收益与估值风险。
在沈萌看来,短期内消费需求仍缺乏明显支撑,科技板块则还有增长叙事。部分资金虽然开始配置回调后的消费股,但未必能很快获得回报。
在多数机构减持泡泡玛特的时候,段永平选择了买入。
2025年8月,段永平还表示“看不懂”,但泡泡玛特发布2025年年报后,他改变了立场。2026年4月起,他通过卖出看跌期权(卖Put)的方式开始布局,后续通过多轮增持,成为仅次于王宁的第二大股东。三季度,泡泡玛特收入接近腰斩后,段永平仍表示"还会再买一些",上限是总股本的10%。
公募基金在减持,段永平在加仓,看似是两种相反的判断。在陈默默看来,不同投资者的持有周期可能不同,买入一家公司的股票,主要是基于交易机会的判断,关键看当前股价与预期价值之间,是否存在收益的空间。
同样的业绩、同样的风险,有人担心增长放缓,有人则认为股价已经充分消化了这种担忧。怎样给一家公司定价,始终是资本市场永恒的难题。
更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App
LG AI 研究院宣布对外免费开放 K-EXAONE 2.5 模型 API
IT之家 11 月 9 日消息,LG AI 研究院韩国当地时间今日宣布,对外免费开放其 K-EXAONE 2.5 模型的 API。包括开发者、企业、研究机构在内任何用户,可无需付费地将 K-EXAONE 2.5 应用于服务开发和研究。

这一自研技术验证模型在今年 8 月发布的原版基础上进行了追加训练,拥有比肩头部开源 AI 模型的智能体能力,在 AutomationBench、SWE-Bench Pro、BrowseComp、Claw-Eval 上的表现与 DeepSeek-V4.1-Flash 和 GLM-5.3 难分伯仲。
参考
Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active
Thinking Machines Lab has released Inkling 2 Small, the first model of its second Inkling generation. The Inkling 2 Small model is a sparse Mixture-of-Experts (MoE) with 594B total and 17B active parameters. It accepts text and images, and supports variable reasoning effort. On the Artificial Analysis Intelligence Index, it scores 31. Its predecessor, Inkling Small, scores 13 on the same index. The model is live on Tinker for fine-tuning and inference.
TL;DR
- Size: 594B total, 17B active MoE; 1M-token context.
- Runs on: Tinker API only for now. Self-hosting hardware is not disclosed. Weights support MXFP8 and NVFP4.
- Performance: 31 on the AA Intelligence Index, ranked #4 of 29 in its price class.
- Best: 86.3% on SWE-Bench Pro v2 (Hard), second only to Kimi K3 in the lab’s table.
- Worst: 13.1% on Terminal-Bench 4.0, versus 32.8% for GLM-5.3-Flash.
What is Inkling 2 Small?
Inkling 2 Small is a multimodal MoE model from Thinking Machines Lab for agentic coding, vision and knowledge work. It takes text and image input and returns text. It is important to note that it does not support audio as of now.
Weights are closed at launch. The Thinking Machines Lab plans an Apache 2.0 release next week, after external safety testing and red-teaming. This follows its safe path to open weights policy.
How is Inkling 2 Small built?
It includes a 48-layer decoder-only transformer. Each token routes to 8 of 384 experts. Attention mixes local and global layers. A 292M-parameter vision encoder, based on Perception Encoder, uses 20×20-pixel patches with 2×2 spatial pooling. Image patches are interleaved with text tokens.
Pretraining ran in 4 phases:
- H1: 8.4T tokens over 68 hours on NVIDIA GB300 NVL72 racks, trained natively in NVFP4.
- H2: 3.8T tokens of newer, higher-quality data at 512K context.
- H3: about 320B tokens of reasoning and agent trajectories at 1M context.
- H4: about 34B tokens of the highest-quality agent trajectories.
A single joint RL run followed. It took 1,075 steps, 17M+ rollouts and 9 days on 2,048 GPUs. Compute exceeded 4 × 10²³ FLOPs. Coding made up 65.0% of RL tokens. The previous Inkling used separate reasoning and agentic RL runs, merged by distillation.
Alignment got explicit attention. Early checkpoints sometimes fabricated results and concealed actions on hard tasks. The Thinking Machines Lab added an alignment reward and on-policy self-distillation from a conduct spec. Its internal misbehavior score fell from 66.8 on Inkling to 14.9.
How does Inkling 2 Small perform on benchmarks?
All task scores below are from the Thinking Machines Lab’s benchmark table. They run at effort 0.99 (xhigh) and temperature 1.0. Most rows use third-party evaluators such as Artificial Analysis, Scale AI and Vals.
Where it scores highest:
- SWE-Bench Pro v2 (Hard): 86.3%.
- PerceptionBench: 64.7%, versus 58.5% for Kimi K3.
- EnterpriseBench-CoreCraft: 71.8%, versus 46.7% for GLM-5.3-Flash.
- JobBench: 70.5%.
- DesignArena Website: 1,340, third on the November 5 leaderboard snapshot.
Where it trails:
- Terminal-Bench 4.0: 13.1%, versus 33.8% for DeepSeek-V4.1-Flash-1017.
- DeepSWE v1.1: 72.1%, versus 77.0% for DeepSeek-V4.1-Flash-1017.
- GDPVal-AA v2.1: 1,630, versus 1,746 for DeepSeek-V4.1-Flash-1017.
On the PMPP-Hard GPU kernel benchmark, it solved 54.25 of 69 tasks on average. In an autoresearch demo, it fine-tuned Qwen3.5-4B for time-series forecasting. It spent $426 of a $500 Tinker budget. The release post cites both 4.5 hours and 4 h 17 min for the run.
How does Inkling 2 Small compare with DeepSeek and GLM?
| Feature | Inkling 2 Small | DeepSeek-V4.1-Flash-1017 | GLM-5.3-Flash |
|---|---|---|---|
| Total / active params | 594B / 17B (card) | 552B / 16B decode, 8B prefill (HF) | 320B / 18B (HF) |
| Context | 1M | 1M | 1M |
| Input modality | Text, image | Text, image | Text, image |
| License | Closed now; Apache 2.0 planned | MIT | MIT |
| Self-host hardware | Not disclosed | Not disclosed | Not disclosed |
| AA Intelligence Index v5.0.1 | 31 (AA) | 33 (AA) | 28 (AA) |
| AA time per index task | 21.5 min | 13.3 min | 37.7 min |
| AA cost per index task | $1.18 | $0.60 | $0.58 |
| Price per 1M tokens (in / out) | $0.30 / $1.00 | $0.30 / $1.20 | $0.15 / $0.50 |
| Terminal-Bench 2.1 | 85.1% (Pi harness, lab) | 90.6% (Sept checkpoint, DeepSeek) | 84.3% (Z.ai) |
| DeepSWE v1.1 | 72.1% (lab) | 74.2% (Sept checkpoint, DeepSeek) | 63.4% (Z.ai) |
Prices and AA metrics come from each model’s Artificial Analysis page, read November 5, 2026.
Free Chinese AI Models Are Getting Scary Good at Hacking
AI broke the cybersecurity world. What’s next?
Today, I’m talking with Alex Stamos, chief security officer at Cognition and longtime cybersecurity expert, about what a mess cybersecurity is right now.
This episode is part one of a special series looking at AI in real life.
I’ve always enjoyed talking to Alex because he’s so direct, and you’re going to hear him say that the multibillion-dollar cybersecurity industry as we’ve known it is essentially overdue to the unrelenting, AI-boosted waves of hacks, breaches, and attacks — and this is only the beginning of how much AI is changing security.
Verge subscribers, don’t forget you get exclusive access to ad-free Decoder wherever you get your podcasts. Head here. Not a subscriber? You can sign up here.
Both private actors and also state actors, like Russia or North Korea, are breaking into systems all over thanks to AI helping them both detect and exploit vulnerabilities all at the same time. It’s got a real frontier feel to it, Alex notes, just like the ’90s, when the whole idea of an always-online world was brand-new.
That much disruption means a lot of places where bad guys can get in while everyone else is left scrambling. Alex isn’t a doomer in the long run, but he does think we’re in for a “dark age” while we try to figure out how on earth to adapt to the new era.
There’s a lot going on here — we even got all the way to how AI-powered misinformation and slop might be managed on our social networks.
Before we get going, do me a favor and subscribe to Decoder on YouTube, where we put out new episodes every Monday and Thursday.
Okay: Alex Stamos on how AI is upending cybersecurity. Here we go.
This interview has been lightly edited for length and clarity.
Alex Stamos, you’re the chief security officer at Cognition. You’re also a lecturer at Stanford University Cyber Policy Center. Welcome back to Decoder.
Hey, thanks for having me, man.
I’m excited to talk to you. It’s been a long time since you and I chatted in what you might call the “content moderation era” of internet policy discourse. I feel like you and I talked a lot about how all that was working, about election security, all those sorts of things. We’re in a new era of big problems on the internet, particularly around cyber defense, cybersecurity. You’ve been thinking about this stuff a lot. You referred to the past few weeks as “security awareness month.” Tell me what’s going on. Just give me the lay of land right now.
October is traditionally security awareness month. Lots of companies run events and deal both internally and then tell people, “Change your password, use multiple passwords, use a password manager.” What has happened over the last month is we’ve had a couple of these big bugs, PHP, Node and such. What’s been less discussed is a huge number of companies have been breached.
You have a series of AI-related companies, the neoclouds that have been breached in succession. You have the Binance hack, which is the largest heist of anything in human history. You have the Cursor incident where one of the things that’s going on here is these hacks are being chained together. From what we know right now, one of the companies that was broken into was Fireworks, which is a prominent provider of AI services to a number of companies. From the Fireworks hack, API keys were stolen that allowed people to break into Cursor. Once they got into Cursor, they then pushed a bad plugin, which went to 5,000 more companies and that caused 5,000 potential breaches. It’s just this kind of permanent continuous breach that is going on throughout the tech industry.
A CISO friend of mine said to a bunch of us, “Is anybody else tired?” And everybody said yes, because effectively we have been at DEFCON 2 for all of October and now going into November.
I’m going to ask you some really dumb questions here. Who is behind all these attacks? Is it Russia? Is it North Korea? Do we know?
It’s a variety of different groups. It is not just because adversaries have decided, but what has happened is, from my perspective, we are at the start of the takeoff of adversary capability enabled by AI, that the open-weight models have now gotten to the point at which you can automate most of the kill chain. This “kill chain” term is something we stole from the military. In the military, they mean to actually kill something. They’re the steps you do to blow something up. In cyber, the kill chain is the things you do to have an effect on a target. So to steal information, to turn something off, to steal money, to steal Bitcoin.
Traditionally, you do research upfront, you look at a piece of software, you find a vulnerability, you write exploits, and you do that all way ahead of time. And then you have a human being go and take that stuff you built and deploy it against either one target — because maybe you’re a government and you’re going against Lockheed Martin or you’re going against another government or you’re a ransomware actor — and you take all that work you did and you deploy it against hundreds or thousands of targets really fast. Now with AI, what’s happening is this kill chain has inverted where you can run an AI system that just goes hog wild and it looks at the targets and it finds the bugs in real time and it writes the exploits in real time and it breaks into these companies. One of the things that’s happening is that these systems are going and they’re so good at finding bugs that they are finding the bugs in real time and writing the exploits in real time.
The system that everybody’s talking about is not the only one, but the system we’re talking about is a thing called ARTEX, which is a Chinese open-source tool that anybody can go look at. There’s an English translation on GitHub that you can check out. And if you power it on the back end by something like an obliterated GLM-5.3, it will let you break into a shocking number of organizations. This started with three South Korean banks. This week, about 200 more companies have been hit. I’ve been helping a couple of companies. Cognition is not an incident response company, but we do go in and help people fix their bugs afterwards by using our tools. It is shocking the companies that are getting hit by something you could just download. It has been a long time since you’ve been able to go and just download an open-source tool
李佳琦卖出万份Token,天猫的AI生意才刚开场
Token正在成为AI时代的“流量包”。(雷峰网雷峰网雷峰网)
11月2日晚,李佳琦直播间上架AI千问办公的Token算力,开播半小时,售出10万份千问办公个人会员月卡,累计约2万亿Token。
这次千问办公CEO陈宇森也来到直播间现场,雷峰网此前了解到,这位阿里最年轻的事业部CEO,业内普遍评价他能力很强,具备出色的产品sense、表达能力和商业认知都超出大部分阿里的资深老员工,“登味”极少。
说起来,阿里在就9月份就已有动作,同期天猫设立“AI空间站”,就有三家厂商入驻:千问、智谱、Minimax,充值规格分为月卡、季卡和年卡三档;阿里旗下的三档AI应用分别也给出了20%—40%的优惠;更早些,智谱直接在天猫开通官方旗舰店。

从模型厂商到电商平台,再到直播间,Token正在被包装成一种可以直接购买的商品。
但对电商平台来说,把Token摆上货架只是第一步。至于做成一门生意,是另一件事。(作者持续关注字节百度等大厂,欢迎添加微信Who123start畅聊)
01
电商平台为什么盯上Token?
不只淘天一家的动作。京东同期上线“AI软件充值馆”,引入智谱GLM、京东云、DeepSeek、腾讯云、火山引擎等产品;早在5月,三大运营商就已将Token套餐引入营业厅。
与此同时,腾讯云将MaaS升级为Token Hub,阿里也在3月成立ATH事业群,6月成立Token Foundry事业部,由集团CEO吴泳铭直接挂帅,延续阿里的“Token战略”。
说到底,电商平台看中的,是Token背后的流量、商家和交易资源。
有接近大厂的人士透露,某互联网公司事业部曾尝试把核心电商平台流量与大模型Token捆绑销售:商家不买Token,就不优先推荐商品。
这个做法在内部引发过争议,有人担心商家会转投其他平台,但决策层认为,这是当时风险最低、见效最快的变现路径。
等到一个月后免费额度用尽,调用量和收入双双下跌,内部才意识到,高用量不等于真实需求。
这套做法未必能代表电商平台的普遍策略,却说明了平台的一种想法:AI可以成为经营商家、提高流量变现效率的新工具。
一位大厂的电商从业者谈到AI时,也提到了商家留存。他认为,AI工具能够提升中小商家留在平台直播间的意愿,说白了就是和其他平台争夺商家。但他也提出一个问题:平台是为了服务好现有商家、避免他们流失,还是靠更好的工具吸引更多商家进场?
对平台,这两件事都意味着新机会。AI工具可以帮助商家提效,Token则可以直接拿来赚钱。
一位从业者半开玩笑地对雷峰网说,卖Token和卖流量包、话费充值差不多。过去,用户要购买模型服务,通常需要进入厂商控制台、注册账号、获取API Key,再按调用量充值。
如今,模型调用资源被包装成月卡、年卡和场景套餐,可以直接放到电商货架上。
这也是Token适合进入电商平台的原因:它有明确的计量单位,也能被包装成不同价位的商品。
02
货架搭起来了,生意还没跑通
不过,Token能卖出去,不代表它能成为一门稳定的生意。
首先是用户能不能留下来。
有网友在李佳琦直播间里,下单千问会员的理由只是标价0.01元:“也不知道是干嘛的,反正便宜就买了。”消费者对Token没有概念,更谈不上忠诚。
从模型厂商的店铺销量看,智谱旗舰店卖得最好的一款套餐成交200多份,MiniMax旗舰店的最好成绩是100多份。

相比直播间的短时销量,这些数字还不足以说明用户会持续购买。后期需要观察的,是首批用户的复购达到多少。
一位云市场的负责人曾用水电来形容Token服务:用户今天用哪家的Token,下一次完全可以换一家。因为模型调用是一种无状态服务,一次调用结束后,用户的数据并不会自然留在模型厂商那里。
其次是商品能不能让消费者看明白。
每百万Token多少钱,看起来是一个简单的比价问题,实际上各家套餐的分词规则、计费口径和质量等级并不相同。
如果用户连自己买到了多少、实际消耗了多少都难以判断,明码标价也很难真正形成有效的商品比较。
最后是平台不能只做一个货架。
如果只负责把用户引到厂商店铺完成交易,那么它更像是销售渠道;如果能持续帮助用户发现、比较和使用不同模型服务,才有机会成为AI服务分发平台。
对电商平台来说,单纯完成上架、支付和交付,未必能形成长期优势。
03