Meta Muse 用量超预期，免费额度撑到习惯养成

Mike @MikeLongTerm原推发布于 2026.10.08

在 X 上打开原推

BREAKING $AMD | $META @Muse is being used harder than what Meta planned 🚀🧵 Muse is already being used harder than Meta planned, and the product gets more expensive to serve as it gets more useful. Alexandr Wang said real users were burning about ten times the tokens of the internal test cohort within a day of launch. That was not a chatbot spike. Each account sits on its own Linux VM, with a browser, files, credentials, and a Sentinel process beside it. More tokens usually means more page reads, screenshots, tool calls, and sub agents, which means the machine is awake and doing work rather than parked. A reported week of ordinary daily use at 34–35 million tokens is still inside the free 100 million allowance, so the median heavy user has not even hit the paywall. Power and Maximum exist for people who want the agent on most of the day. The cap is generous enough that habit can form before Meta starts charging for it. That habit is what pulls CPUs. A normal cloud service multiplexes idle users down to almost nothing. Muse cannot. The security model is one isolated computer per person, and the product promise is that work continues after the phone is closed. Idle accounts can be suspended, and waiting agents can share cores, but a user running a browser loop for a few hours a day still needs a resident VM with gigabytes of memory. If usefulness rises, more connectors, fewer failed tasks, cron jobs that people actually trust, the live fraction of daily users rises with it. The early strain below a million daily users, including degraded service and sub agent spawns that failed under load, is the same mechanism at small scale. Token growth and hour growth are not the same meter, but both point the same way: more concurrent sandboxes. That sandbox layer is an @AMD server CPU problem. Observed Muse hosts are EPYC class parts, and the next dense part, Venice, tops out at 256 cores. Meta can pack many VMs on a chip and oversubscribe threads that are waiting on the network. It is still a purchase cycle Meta does not have for a stateless model API. Every step up in hours per user, or in the share of users who leave the agent running, raises the number of live VMs Meta has to keep warm. Those VMs land on EPYC trays. If Muse moves from a few million weekly users toward tens or hundreds of millions, the incremental CPU order is not a rounding error in AMD’s server year, even after hibernation and oversubscription. The bullish thing is that usage intensity is already ahead of plan, the architecture converts intensity into resident cores, and AMD is the vendor on the host. Not Financial Advice! DYOR!

这条是引用着另一条写的：「Muse 每个用户一台 Linux 虚拟机，CPU 需求暴涨」