Meta 发布了 Muse Wonder，内部代号为 Watermelon，并称其在编码和专业工作方面与 ChatGPT 和 Claude 相当。它尚未说明是否会公开权重，而这决定了它在欧盟《人工智能法案》下应承担的责任。
Meta has released Muse Wonder, known internally as Watermelon, and says it matches ChatGPT and Claude on coding and professional work. It has not said whether the weights will be published, which decides what it owes under the EU AI Act.
Meta 发布了 Muse Wonder，这是 AI 部门在马克·扎克伯格花费十四个月重建后推出的最强大模型。内部它被称为 Watermelon。
Meta has released Muse Wonder, the most capable model to come out of the AI unit Mark Zuckerberg spent fourteen months rebuilding. Internally it was Watermelon.
首席人工智能官亚历山大·王告诉彭博社，该模型在编码以及金融和法律工作方面与 ChatGPT 和 Claude 达到同等水平。该说法未附带公开的基准测试结果。
It claims parity with ChatGPT and Claude on coding and on financial and legal work, chief AI officer Alexandr Wang told Bloomberg. The claim arrives without published benchmarks.
Meta 达到了其内部时间表。The Information 在八月报道称，公司正瞄准十月发布 Watermelon，但未说明是否会以 Muse 名称推出。
Meta has hit its own internal date. The Information reported in August that the company was targeting October for Watermelon, without establishing whether it would ship under the Muse name.
这是今年第三次提出前沿模型的声明。王在九月对 Muse Spark 1.3 也作了类似表述，而扎克伯格在四月曾称第一代 Muse Spark 为世界级助手。
It is the third frontier claim this year. Wang said much the same about Muse Spark 1.3 in September, and Zuckerberg called the first Muse Spark a world-class assistant in April.
第一代模型附带了一份评分卡。Meta 自身的数据显示，其在某一指数上排名第四，得分为 52，而领先者得分为 57；在两项推理基准测试中则落后于他人。
That first one came with a scorecard. Meta’s own figures placed it fourth on one index, at 52 against 57 for the leaders, and behind on two reasoning benchmarks.
面对质疑，王的回答是让持怀疑态度的人亲自尝试并自行判断。基准测试本身证明力有限，因为它们可能被操纵，且会因任务而异。
Wang’s answer to doubters is that they should try it and judge for themselves. Benchmarks are a weak proof anyway, since they can be gamed and vary by task.
实际发布的范围较窄。开发者可通过 Meta Model API 和 Muse Code 访问该模型，且它将为 Meta 上月推出的 Muse 代理的付费层提供支持。
What has actually shipped is narrow. Developers can reach it through the Meta Model API and Muse Code, and it will power the paid tier of the Muse agent Meta launched last month.
其背后的投入却并不窄。扎克伯格已承诺投入数千亿美元用于基础设施建设，并花费 140 亿美元收购 Scale AI 的股份，主要目的是招聘王。股价一度上涨了 4.7%。
The spending behind it is not narrow. Zuckerberg has committed hundreds of billions to infrastructure and paid $14B for a stake in Scale AI largely to recruit Wang. Shares rose as much as 4.7%.
对欧洲而言，问题不在于基准测试。而在于 Meta 是否会公开权重，以及其记录是否按规模划分——对提炼后的 30B 模型持开放态度，对旗舰模型则保持封闭。
For Europe the question is not the benchmark. It is whether Meta publishes the weights, and its record splits by size, open for a distilled 30B model, closed for the flagship.
根据《人工智能法案》第53条，真正免费且开放的模型可以免除部分向人工智能办公室提交的文档义务。许可证必须是真实的，模型权重、架构和使用信息需公开。
Under Article 53 of the AI Act, a genuinely free and open model escapes some of the documentation owed to the AI Office. The licence has to be real, with weights, architecture and usage information public.
Meta尚未说明Muse Wonder将采取何种方式。此前，它也尚未决定Muse Spark 1.3的处理方式，而扎克伯格已表明对开放最强大系统持谨慎态度。
Meta has not said which way Muse Wonder goes. It had not decided for Muse Spark 1.3 either, and Zuckerberg has signalled caution about opening the most powerful systems.
不过，该豁免仍有上限。无论许可证如何，存在系统性风险的模型必须承担全部义务，而Meta拒绝签署为落实这些义务而制定的行为准则。
The exemption has a ceiling regardless. Models carrying systemic risk owe the full set of obligations whatever their licence, and Meta declined to sign the code of practice written to operationalise them.
此事具有双刃剑效应。Meta在Muse Image上线三天后将其叫停，原因是其默认设置被发现会自动选择公开账户。
Speed cuts both ways here. Meta killed Muse Image three days after launch when its default settings turned out to opt public accounts in.