去年,扎克伯格「千金买马骨」,挥舞支票满世界招揽 AI 人才,部分薪酬思路高达上亿美元,手笔之大,堪比 NBA 球队砸顶薪抢巨星。如今,高价招揽人才、重组团队和扩建算力的投入,终于有了一份更接近预期的成绩单。
就在刚刚,Meta 发布 Muse Wonder 和 Muse Wonder Full。两款模型此前代号为 Watermelon,主攻个人与专业智能体任务,重点提高跨对话记忆、长时间执行任务和编程能力。
扎克伯格在接受采访时表示,Meta 已经抵达 AI 工艺的最前沿。

按照 Meta 公布的实测,Muse Wonder Full 新模型至少在部分基准测试上,已能与 GPT-6.1 Astra 和 Claude Fable 5.1 掰手腕。第三方机构 Artificial Analysis 的测试也表明,Wonder 系列较上一代脱胎换骨,在多项智能体和专业工作实测中跻身前三甲。

花大价钱攒起来的队伍,让 Meta 重新有了跟领先模型叫板的底气,也让 Meta Muse 的 Personal Agent 之路能走得更顺畅些。
更强的记忆和执行力,能否让 AI 善始善终
Meta 此次发布,延续了一年多来对 AI 业务的重整。
2025 年 4 月,Llama 4 马失前蹄,扎克伯格随即加大人才和算力投入,招揽 Scale AI 创始人 Alexandr Wang,重组研究团队。今年 4 月,重组后的团队推出 Muse Spark,但与竞争对手相比仍差着一口气。
按照 Meta 的说法,过去 14 个月里,团队推倒重来,从底层重建工艺体系,并持续扩大预训练和强化学习的规模。此次 Wonder 系列在吉瓦级 Prometheus 数据中心完成训练,成为这轮豪赌交出的最新筹码。
扎克伯格想回答的难题也很简单:AI 已经能解奥数级的数学题,却未必接得住一通客服电话;一个理想的个人助理,得有应付柴米油盐的常识和执行力。

Muse Wonder 围绕两个方向动刀,分别是 Active Memory 和 Active Focus。
Active Memory 负责跨对话保留关键信息,并在新任务需要时主动找回相关内容。使用者此前提过的偏好、限制和个人情况,可以成为后续任务的参考,减少每次对话都重新交代背景的麻烦。
Wang 向 Axios 表示,个人智能体需要让关键信息在数天、数周乃至数月的交流中延续下来。





向左滑动查看更多内容
Active Focus 则针对任务执行到一半停下来的难题。现实中的操作会遇到页面变化、信息缺失或预订失败,模型需要尝试其他方法,依据新信息调整计划,并持续推进任务。Meta 表示,模型仍会在必要时询问使用者,并在关键操作前征求确认。
扎克伯格在最新访谈中将这一思路概括为「模型就是机型」,能力和性格等机型特征都得进训练过程。Wang 补充,智能体何时主动、如何与使用者交流,也属于需要专门外观的机型行为。

Wang 向 CNBC 举了预订餐厅的例子。智能体查询空位时可能遭遇各种障碍,训练的目标之一,就是让它尝试其他路径,减少一遇到难题便停下来向使用者求助的情况。
编程是检验这些能力的试金石。
一个持续开发的软件项目,要求模型做到心中有数:记住需要、读懂已有代码、改动多个文件、反复测试修正。Meta 表示,两款模型都接受过多种长时间任务的训练,以提高持续作战的能力。

官方展示还涵盖「看图做菜」,依据图片生成可玩的 3D 厨房游戏,以及结合个人偏好、调动浏览器和多个应用安排跑步计划。
面向专业工作,Meta 列出了法律、金融和健康相关实测,其中 Harvey 的 Legal Agent Benchmark 要求模型依据原始材料和详细指令生成法律工作成果。

Artificial Analysis 的实测表明,Wonder 系列综合能力显著提高,Full 版十项测试有六项杀入前三。在最高推理强度下,Muse Code 搭配 Wonder Full 的编程智能体成绩与 Claude Code 搭配 Sonnet 5.5 平分秋色,还略胜搭配 Opus 5.5 一筹。

社交平台上的早期感受,则是有人欢喜有人愁。
使用者 Chetaslua 展示了一段包含《龙珠》角色的动画,称画面完全由代码生成,并对 Muse 的响应速度竖起大拇指。他同时注明,配音来自 ElevenLabs,音乐由 Suno 制作。


另一位使用者 Max Blade 则给 Muse Wonder Full 和 Opus 5.5 出了同一道题:制作一款可以边喝啤酒、边抽雪茄的「3A 级割草机模拟器」。

从他发布的比较画面看,Opus 版本的草地、树木和光影纤毫毕现,Muse 版本的情形与人物手部则显著粗枝大叶。他认为,Muse 在此类 3D 游戏任务上仍欠火候。
使用者 Bhavy 分享的火箭情形比较也有类似反馈。他表示,Muse Wonder 的成效不如 Kimi K3,自己记录的花费却达到 14.3 美元,而 Kimi 约为 6.9 美元,堪称事倍功半。从配图看,Muse 的情形更偏卡通,Kimi 版本则展示了更多火箭结构与烟雾细节。

Meta AI 负责人 Alexandr Wang 在社交平台回应使用者时也承认,新模型的编程能力还不如 Opus 5.5,团队正在排查多个 bug。

对于部分使用者不满意的 3D 游戏演示成效,Wang 解释,团队没有专门针对这类演示开展训练,希望使用者把模型用到真刀真枪的编程工作中。他表示,团队依据 Muse 使用者的实际需要挑选训练方向,同时承认可以在演示情形上多下功夫。

Meta 在官方博客中还试图回答另一个难题:模型的进步能否举一反三,迁移到训练时没有重点覆盖的任务。
团队比较了 Muse Spark 1.3 与 Muse Wonder Full 在「熟题」和「生题」上的表现。按照 Meta 的解释,随着推理计算量加码,两款模型都在进步,但 Wonder Full 在陌生任务上的表现,仍高出按 Spark 趋势推演的预期一截。

Meta 将其归因于底子厚:更大、更强的基础模型,让后训练的红利可以触类旁通。
对于需要处理不同目标、工具和限制的个人智能体,这种迁移力尤其关键。不过,这部分结论出自 Meta 的「自家答卷」,具体能覆盖哪些真实情形,还得让子弹再飞一会儿。
榜单之外,Meta 还要说服使用者长期掏腰包
模型能力升级的同时,Meta 也在趁热打铁推动收费。
发布时,Muse Wonder 已借助 Meta Model API 和 Muse Code 开放,能力更强的 Wonder Full 则仅向 Muse Code 订阅使用者给出。依据 CNBC 报道,Muse Code 的相关订阅档位为每月 50 美元和 200 美元。
普通使用者使用的 Muse 个人助理应用,将在后续接入 Wonder,并优先向订阅使用者开放。
Muse 没有模型挑选器,也没有推理强度选项。Wang 认为,普通使用者不该为了用 AI 而啃工艺说明书,付费升级后也不必手动挑模型,讲究一个开箱即用。
在 API 维度,Muse Wonder 售价为每百万输入 token 4 美元、每百万缓存输入 token 0.20 美元,以及每百万输出 token 20 美元。两款模型均兼容 100 万 token 上下文,可接收文本、图像、视频和音频输入,输出形式为文本。

而按照 Artificial Analysis 的 Intelligence Index 测试,Muse Wonder 在最高推理强度下,每项任务成本约为 4.52 美元,低于同次比较中的 Claude Fable 5.1、Opus 5.5 和 Sonnet 5.5,却高于 GPT-6 Astra 的 3.26 美元、Gemini 4 Argon 的 1.99 美元,以及 GPT-6.1 Sol 的 0.72 美元。
Artificial Analysis 指出,Wonder 虽然比上一代省输出 token,但较高的输入消耗与 API 定价仍然推高了总账。使用者最终要算的,还涵盖完成同一项工作要烧多少 token,以及中途要返工几次。
订阅额度的消耗也引来了一些使用者的不满。
在前述割草机模拟器测试中,Max Blade 为 Muse Wonder Full 和 Opus 5.5 标注的「API 等效成本」分别为 4.11 美元和 99.43 美元,但他称,Muse 胃口太大,已耗尽自己 50 美元套餐的可用额度,迫使他升级到 200 美元套餐才跑完任务。

当然,作为开源模型的老大哥,此次发布也让 Meta 的「开源人设」再次受到审视。Wang 表述 Meta 初心不改,仍然兼容开源,并提到此前已开放 Muse Glimmer 的权重,以及将 Muse 智能体部署到硬件上的相关工艺。
相比此前以开放换朋友圈的路线,Wonder 让 Meta 更直接地下场,在订阅、API 和企业客户市面贴身肉搏。CNBC 指出,Meta 正尝试在核心广告业务之外,再开一条财路。
但智能体能否长成摇钱树,还取决于模型之外的变量。
最近,一些使用者在社交平台投诉 Muse 未经授权访问个人消息,Meta 对此则提出异议,称部分情况可能是使用者在较早的对话中授过权,转头忘了;团队正在跟踪这些难题,并协助使用者弄清授权是如何生效的。

双方各执一词,但 Personal Agent 的机型外观必须面对一个现实:使用者可能在很久之前点过头,模型却会在后续任务中拿旧账当令箭,授权范围和有效时间因而需要一目了然。
Meta 表示,Wonder 系列加强了抵御恶意指令、处理私人信息,以及防范网络和生物领域滥用的能力,后续还将发布更完整的安全报告。
扎克伯格则提到尚未推出的 Muse Confidential VM,称该思路计划让使用者数据和与智能体的对话对 Meta 自身也密不透风。
有趣的是,在接受《金融时报》采访时,扎克伯格拒绝了给 AI 发展「踩刹车」的主张,认为各家实验室应借助工艺措施和负责任的开发来保障安全。他还透露,Meta 已在研发下一阶段的模型,并预计之后的进展会更快。
持续提高的模型能力,最终还是要支撑起扎克伯格设想中的 Personal Agent,让 AI 可以处理个人日常事务。用最前沿的工艺订餐、安排行程、处理邮件,听起来有些大材小用,但这些任务都要面对不断变化的信息和各不相同的个人要求。
Muse Wonder 挑选的方向,也让模型进步有了更具体的衡量方式。AI 的能力可以由最难的题目来检验,个人智能体的价值,则要由那些每天重复、又总得有人操心的小事来证明。
本文来源:https://www.ifanr.com/1683963
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Last year, Zuckerberg "bought money and bones", wielding checks all over the world to recruit AI talents, and some pay ideas were as large as hundreds of millions of dollars, which is comparable to NBA teams hitting the top salary and grabbing superstars. Now, the investment of high-priced recruiting, restructuring teams and expanding computing power finally has a report card that is closer to expectations.
Just now, Meta released Muse Wonder and Muse Wonder Full. The two models, formerly codenamed Watermelon, focus on personal and professional agent tasks, focusing on improving cross-dialog memory, long-term task execution and programming capabilities.
Zuckerberg said in an interview that Meta has reached the forefront of AI technology.

According to the actual measurement published by Meta, the new model of Muse Wonder Full has been able to break the wrist with GPT-6.1 Astra and Claude Fable 5.1 at least on some benchmark tests. Testing by third-party agency Artificial Analysis also showed that the Wonder series was reborn from the previous generation and ranked among the top three in a number of agent and professional work tests.
Spending a lot of money to save up the team, so that Meta has the courage to follow the leading model again, but also so that Meta Muse's Personal Agent road can go smoother.

The release of Meta follows more than a year of restructuring of the AI business.
In April 2025, Llama 4 lost his front foot, and Zuckerberg immediately increased his talent and computing power, recruiting Alexandr Wang, founder of Scale AI, and restructuring the research team. In April this year, the restructured team launched Muse Spark, but it is still a bit of a sigh of relief compared to its competitors.

According to Meta, over the past 14 months, the team has rebuilt the process system from the ground up and continued to scale up pre-training and reinforcement learning. The Wonder series completed training in the Gigawatt Prometheus data center, becoming the latest chip in this giant round of gambling.
The puzzle Zuckerberg wants to answer is also simple: AI can solve mathematical problems, but it may not be able to accept a customer service call; an ideal personal assistant must have the common sense and execution ability to deal with diesel oil and salt.
Muse Wonder moves the knife in two directions, Active Memory and Active Focus.

Active Memory is responsible for retaining key information across conversations and proactively retrieving new tasks as needed. The user's previously mentioned preferences, limitations, and personal circumstances can serve as a reference for subsequent tasks, reducing the hassle of reconfirming the context of each conversation.
Wang told Axios that personal agents need to allow key information to persist over days, weeks and even months of communication.

Swipe left for more
Active Focus addresses the puzzle of halting halfway through a task. The actual operation will encounter page changes, missing information, or unsuccessful reservations, and the model needs to try other methods, adjust the plan according to the new information, and continue to advance the task. Meta says the model will still ask the user if necessary and ask for confirmation before critical operations.

In his latest interview, Zuckerberg summed up this line of thinking as “a model is a model”, and model characteristics such as ability and personality have to go into the training process. Wang added that when the agent takes the initiative and how to communicate with the user is also a model behavior that requires a special appearance.
Wang gave CNBC an example of a restaurant reservation. The agent may encounter various obstacles when querying for vacancies. One of the goals of training is to let it try other paths to reduce the number of cases where it stops and asks the user for help as soon as it encounters difficulties.

Programming is the litmus test of these abilities.
A software project in continuous development requires the model to be mindful: remembering needs, reading existing code, changing multiple files, and repeatedly testing corrections. Both models have been trained for a variety of long-duration missions to improve their ability to operate continuously, Meta said.
The official showcase also covers “cooking with pictures”, generating playable 3D kitchen games based on pictures, and arranging running plans based on personal preferences, mobilizing browsers, and multiple apps.

For professional work, Meta lists legal, financial, and health-related measurements, where Harvey's Legal Agent Benchmark requires models to generate legal work based on raw materials and detailed instructions.
The experiments of Artificial Analysis show that the comprehensive ability of the Wonder series has been significantly improved, and six of the ten tests in the full version have been in the top three. At the highest inference strength, Muse Code with Wonder Full's programming agent scores are evenly divided with Claude Code with Sonnet 5.5 and slightly better with Opus 5.5.

Some people are happy and some are worried about early feelings on social platforms.
User Chetaslua shows an animation featuring a Dragon Ball character, saying the footage is generated entirely by code and thumbs up to Muse's responsiveness. He also noted that the voiceover was from ElevenLabs and the music was produced by Suno.

Another user, Max Blade, posed the same question to Muse Wonder Full and Opus 5.5: make a “Class 3A lawn mower simulator” that can drink beer and smoke cigars.
From the comparison he released, the Opus version of grass, trees, and light and shadow fibers appeared, while the Muse version and the characters' hands were significantly rough. He believes that Muse is still under fire for such 3D game missions.

User Bhavy shared similar feedback on the comparison of the rocket situation. He said that Muse Wonder was not as effective as Kimi K3, but his own record of spending was $14.3, while Kimi was about $6.90, which was half the success. From the illustration, Muse's situation is more cartoony, and the Kimi version shows more rocket structure and smoke details.
Alexandr Wang, the head of Meta AI, also acknowledged in response to users on the social platform that the new model was not as programmable as Opus 5.5 and that the team was troubleshooting multiple bugs.
Regarding the performance of 3D game demonstrations that some users are dissatisfied with, Wang explained that the team did not specifically train for such demonstrations, and hoped that users would use the model in the programming work of real knives and real guns. He said that the team selected the training direction based on the actual needs of Muse users, while acknowledging that they could work harder in the demonstration situation.

Meta is also trying to answer another conundrum in the official blog: whether model advancements can be replicated by migrating to tasks that are not covered by the focus at the time of training.
The team compared Muse Spark 1.3 with Muse Wonder Full on Problems and Problems. As Meta explains, both models are improving as inference calculations increase, but Wonder Full's performance on unfamiliar tasks is still higher than expected based on Spark trends.

Meta attributes it to a thick base: a larger, stronger base model that allows post-training dividends to be bypassed.
This mobility is especially critical for individual agents that need to deal with different goals, tools, and constraints. However, this part of the conclusion comes from Meta's "own answer sheet", which can cover the real situation, and you have to let the bullet fly for a while.

At the same time as the model's ability is upgraded, Meta is also pushing for charging while the iron is hot.
At launch, Muse Wonder is available via the Meta Model API and Muse Code, while the more capable Wonder Full is only available to Muse Code subscribers. According to CNBC, subscription tiers for Muse Code are $50 and $200 per month.

The Muse personal assistant application used by ordinary users will be connected to Wonder in the future and will be open to subscribers first.
Muse has no model picker and no inference strength option. Wang believes that ordinary users should not chew the process manual in order to use AI, and after paying for upgrades, they do not have to manually pick the model and pay attention to it out of the box.
In the API dimension, Muse Wonder sells for $4 per million input tokens, $0.20 per million cache input tokens, and $20 per million output tokens. Both models are compatible with 1 million token contexts and can receive text, image, video, and audio inputs in the form of text.

According to Artificial Analysis's Intelligence Index test, Muse Wonder costs about $4.52 per task at its highest inference strength, which is lower than Claude Fable 5.1, Opus 5.5, and Sonnet 5.5 in the same comparison, but higher than GPT-6 Astra's $3.26, Gemini 4 Argon's $1.99, and GPT-6.1 Sol's $0.72.
Artificial Analysis pointed out that although Wonder output tokens are lower than the previous generation, higher input consumption and API pricing still push up the general ledger. The user's final calculation also includes how many tokens to burn to complete the same work, and how many times to rework in the middle.

Consumption of subscription credits has also caused dissatisfaction among some users.
In the aforementioned mower simulator tests, Max Blade marked the “API equivalent cost” of Muse Wonder Full and Opus 5.5 at $4.11 and $99.43, respectively, but he said that Muse's appetite was so great that he had exhausted his $50 package, forcing him to upgrade to the $200 package to complete the task.

Of course, as the big brother of the open source model, this release also makes Meta's "open source people" again under scrutiny. Wang stated that Meta is still compatible with open source without changing its mind, and mentioned the weighting of the previously opened Muse Glimmer and the related process of deploying the Muse agent to the hardware.
Compared with the previous route of opening up the circle of friends, Wonder has made Meta more direct underground and intimate in the subscription, API and enterprise customer market. CNBC noted that Meta is trying to carve out a path beyond its core advertising business.

But whether the agent can grow into a cash cow also depends on the variables outside the model.
Recently, some users complained on social platforms about Muse's unauthorized access to personal messages, which Meta disputed, saying that in some cases, the user may have granted authority in an earlier conversation and forgot to turn his head; the team is tracking these difficulties and helping users understand how the authorization works.
Both sides insist on each other, but the appearance of the Personal Agent model must face a reality: the user may have nodded too much a long time ago, but the model will take the old accounting order arrow in the follow-up task, so the scope of authorization and effective time need to be clear at a glance.

Meta said that the Wonder series strengthens the ability to resist malicious commands, process private information, and protect against abuse in the cyber and biological fields, and will be followed by a more complete security report.
Zuckerberg, referring to the yet-to-be-launched Muse Confidential VM, said the idea was to keep user data and conversations with agents private from Meta itself.

Interestingly, in an interview with the Financial Times, Zuckerberg rejected the idea of “hitting the brakes” on AI development, arguing that laboratories should rely on process measures and responsible development to ensure safety. He also revealed that Meta is already working on the next phase of the model and expects progress to be faster.
The continuous improvement of model capabilities will ultimately support Zuckerberg's envisioned Personal Agent, so that AI can handle personal daily affairs. Using cutting-edge processes to order meals, schedule rides, and process emails may sound daunting, but these tasks are subject to ever-changing information and varying individual requirements.

The direction chosen by Muse Wonder also gives a more concrete way to measure the progress of the model. The capabilities of AI can be tested by the most difficult questions, and the value of individual agents is demonstrated by the small things that are repeated every day and that someone has to worry about.
This article is from: https://www.ifanr.com/1683963 For copyright issues, please write to: 13951431913#139.com or submit a review on the right.