Muse Wonder 对比 MiniMax M3：23 个指数点与 13 倍的输出价格，方向截然相反

MiniMax M3 每百万输入 token 收费 $0.30，每百万输出 token 收费 $1.20。Muse Wonder 则分别为 $4.00 和 $20.00。这意味着输出这一项上有 13.3 倍的差距，而在每完成一项任务的实测成本上有 16.7 倍的差距——在同一个独立评测框架上，MiniMax M3 约为 $0.51，而 Muse Wonder 约为 $4.52。在能力方面，同一评测框架下两者相差 23 个指数点，分别为 52 和 29。所以，显而易见的解读在读到第二段之前就已尘埃落定：Meta Superintelligence Labs 于 2026 年 10 月 12 日发布的模型是那个昂贵而强大的模型，而 MiniMax 的开放权重 M3 则是那个便宜且较弱的模型。让这对组合值得写一整页的原因在于，这两个模型对同一个工程问题押下了相反的赌注，而且其中一个公开了一个另一个没有公开的数字。

Muse Wonder 宣称拥有 1,048,576 token 的上下文窗口，并对窗口内的每一个 token 给出完全相同的定价。MiniMax M3 宣称同样的窗口规模，随后又在其下公布了一个保证可用的最低值 512K，因为它的稀疏注意力设计存在一个必须承诺的下限。这两种说法哪一种对你更有价值，完全取决于你的提示词实际落在窗口的哪个位置。

这两个模型，以及为什么它们不是同一类对象

MiniMax M3 is M​iniMax's flagship open-weight foundation model, announced on the vendor's blog on June 1, 2026 and listed in our own catalogue as available since May 31 — a one-day discrepancy of the kind these records always carry. It is a 428-billion-parameter mixture-of-experts model with roughly 23 billion parameters active per token, 60 layers, 128 routed experts with top-4 routing, and a 1,048,576-token context window with a 512,000-token maximum output. It accepts text, image and video input and returns text. Its three advertised claims are frontier-level coding and agentic performance, native multimodality including an ability to operate a desktop computer, and — the one that matters here — MSA, or M​iniMax Sparse Attention, a proprietary attention architecture the company proposed. It is open weights under the M​iniMax Community License, which is not Apache 2.0 and not MIT, and the distinction is worth reading rather than assuming.

Muse Wonder 是 Meta Superintelligence Labs 于 2026 年 10 月 12 日发布的前沿成果，以双检查点系列的形式推出。只有 muse-wonder-1.0 设有端点——位于 Meta Model API 和 Muse Code 内部——此处所有数据均适用于它。得分更高的配置 Muse Wonder Full 仅通过 Muse Code 订阅出售。可调用的检查点接受文本、图像、语音和视频输入，返回文本，拥有 1,048,576 token 的上下文窗口，单次响应上限为 943,718 token，价格为每百万 token 输入 $4.00、缓存 $0.20、输出 $20.00，并且强制启用推理，其努力程度阶梯从 minimal 到 max，默认为 medium。没有权重，也没有许可证。

所以这个组合并不只是廉价与昂贵的对比。它是可下载的 428B 开放权重模型与租用的专有模型之间的较量，这意味着价格比较和能力比较回答的是不同的问题。

23个指数点是由什么构成的

这两组数据都来自Artificial Analysis对同一工具同一版本的测试，且都是在最高努力程度下得出的，这也是它们能够被并列比较的唯一原因。

智能指数 — Muse Wonder 52，在 216 个模型中排名第九；MiniMax M3 29，在其类别 116 个中排名第十八

• 每完成一个索引任务的成本 — MiniMax M3 约 $0.51；Muse Wonder 约 $4.52，约为 8.9 倍

• Output tokens consumed on that run — MiniMax M3 about 120 million; Muse Wonder about 130 million, against a board median near 140 million for M​iniMax's cohort

• 标价 — MiniMax M3 每百万 $0.30 输入 / $0.06 缓存 / $1.20 输出；Muse Wonder $4.00 输入 / $0.20 缓存 / $20.00 输出

• 上下文与输出上限——输入均为 1,048,576；MiniMax M3 输出 512,000，而 Muse Wonder 为 943,718

• 输入模态 — MiniMax M3 文本、图像和视频；Muse Wonder 文本、图像、语音和视频

• Weights — MiniMax M3 downloadable under the M​iniMax Community License; Muse Wonder proprietary, none published

• 速度，经独立测量——MiniMax M3 每秒约输出 108 个 token，首 token 时延不到 1 秒；Muse Wonder 每秒约 115 个 token，首 token 时延接近 20 秒

最后一行改变的是工作负载的体感，而不是它的成本；在价格之后，它是清单上最大的实际差异。对于任何交互式场景来说，亚秒级的首 token 时间和二十秒级的首 token 时间是完全不同的产品，无论每个模型最终开始生成时的得分如何。

另请注意，这两个模型在输出 token 量上很接近——1.2 亿对 1.3 亿——这意味着每任务成本比本质上就是价目表上的比率，而不是某个模型啰嗦造成的假象。当一个便宜模型既在每 token 上便宜，又在使用的 token 数量上克制，那么价格差异就是真实存在的，不会在最终开票的路上被吃掉。

成本算术，以两种至关重要的形态呈现

以两个工作负载为例。第一个是摘要或抽取流水线：每月 1000 万 token，输入占比 70%，这是大多数估算工具默认采用的形态。第二个是长上下文检索：每月 500 万 token，每次调用的提示为 80 万 token，响应很短，且已缓存。

• 短调用量，无缓存——MiniMax M3 每月约 5.70 美元；Muse Wonder 每月约 11.30 美元。大约 2 倍，而对于这种规模的工作负载，与工程师的时薪相比，这两个数字都不过是舍入误差。

• Short-call volume, caching on — MiniMax M3 about $4.86; Muse Wonder about $9.63. The cache line costs M​iniMax five cents per million against M​eta's twenty, but the absolute gap stays small.

• 在 80 万 token 提示词下的长上下文检索——MiniMax M3 每次调用约 0.35 美元；Muse Wonder 每次调用约 4.00 美元，因为这两款模型都不实行阶梯费率，而 Muse Wonder 的输入单价要高出 13 倍。按每月一千次调用计算，这就是 350 美元对 4000 美元。

• 使用缓存时同样的退休工作负载——MiniMax M3 每次调用约 $0.14；Muse Wonder 每次调用约 $2.20

这些数字都不是对你账单的估算——它们只是按各家厂商公布费率进行的标价算术。关键在于其形态：对小型工作负载而言，13倍不过是咖啡钱；而对真正长上下文的工作负载来说，它决定了一项服务你能否负担得起大规模运行，还是只能有选择地分流使用。

MSA 相对于平坦窗口，这是对失效模式的一场押注

这是对比中没有基准行的部分，也是两个模型做出相反选择的地方。

M​iniMax attributes M3's long-context behaviour to MSA, which the company describes as supporting a million tokens with a guaranteed usable minimum of 512K. That word — guaranteed — is doing real work. A sparse-attention system degrades in ways a dense one does not, and publishing a floor is the vendor telling you where it believes the degradation becomes unacceptable. If your prompts sit between 512K and 1M, that is a statement about the risk you are taking, and it is a more useful piece of information than a bare window size.

M​eta's Muse Wonder publishes 1,048,576 tokens and says nothing comparable about a usable floor. It also charges the same rate at 900,000 tokens as at 900, and it caps a single response at 943,718 tokens, which is the second-largest output ceiling we track. What it does not do is tell you where its long-context recall starts to sag. For an agent that opens a large context and works inside it for a long session, the absence of that number is a gap you will discover on your own traffic rather than read on a page.

The honest summary is that M​iniMax has published a guarantee it may or may not keep, and M​eta has published a window with no promise attached to it. Neither statement substitutes for testing your own prompts at your own lengths.

开放权重，以及许可证实际到底说了什么

MiniMax M3 is downloadable, which means three things a rented model cannot offer: you can fine-tune it on proprietary data, you can run it inside your own perimeter, and you can be confident that the artefact answering your calls in six months is the same one answering them today. The M​iniMax Community License is not an unrestricted licence, and anyone planning to build a product on top of it should read the terms rather than assume MIT-style freedoms — the practical question is what it permits for redistribution and for commercial use at scale, and the answer lives in the licence text, not in a blog post. Compare with Apache-2.0 open-weights checkpoints, which is a materially different grant.

Muse Wonder offers none of that. It is API-only through M​eta's own surfaces, which is the trade a frontier proprietary release always makes: no fine-tuning, no self-hosting, no version you control, in exchange for capability that open weights at this price tier do not currently reach. Two days into its life, the independently measured gap is 23 index points, and that is the number you are paying the rent for.

在没有第二份合同的情况下让它们相互较量

The comparison is unusually easy to run here, because one side of it is already on a key you may hold. OrcaRouter carries MiniMax M3 as minimax/minimax-m3 at M​iniMax's own list price with 0% markup — provider list price passed straight through, so a vendor price cut is live on our side the same day it is announced — sitting alongside more than 200 other models behind one credential, with automatic failover across providers when an endpoint is unhealthy. Both of the numbers this page argues about are therefore checkable against your own token counts without a second vendor relationship.

Muse Wonder is not in our catalogue, and there is no third-party serving coverage we can verify. A router cannot route a model nobody else is serving, and the only path to it today is M​eta's own API and Muse Code. Where the routing story does apply is on the boundary: a routing DSL that composes several models into one call is what lets a cheap model carry the volume and a frontier model take the hard requests, and automatic failover keeps a single unhealthy endpoint from taking the whole path down. That pattern needs both models to be reachable through the same interface, which is true for one of these two and not yet for the other.

哪一个，给谁的

当工作负载是高容量且预算受限时，选择 MiniMax M3。以每百万 $0.30 和 $1.20 计，其输入成本约为 Muse Wonder 的十六分之一，输出约为十三分之一，它的首 token 速度快一个数量级，它支持百万 token 窗口并公布了价格下限，而且它是两者中你可以下载、微调并保留的那个。对于规模化摘要、抽取、分类，以及原本仅输入每次调用就要花四美元的长上下文检索，那 23 点的指数差距并不能描述你的工作负载。

当工作艰难、漫长且具有智能体性质时，当你需要的是能力而非价格时，当 943,718 token 的响应上限或强制推理阶梯是一种特性而非负担时，选择 Muse Wonder。在两者之中，它是唯一一个独立测得分数高于 50 的，而且它的定价是三大主要实验室中任何一家为前沿模型公布过的最低公开标价——对于一款成本是 M3 的 13 倍的模型来说，这句话很奇怪，但也很公平。

不该在两者之间做取舍的买家，是流量具有两种形态的那一类。把一份 80 万 token 的提示词放到每百万输出 20 美元的模型上，与把一个仅两段话的分类任务放到它上面，是两套不同的成本模型；而本页上的这两个模型，其定价几乎就像是市场早已心知肚明。