cd /news/ai-agents/meta-muse-gives-you-a-pretty-nice-vm… · home › topics › ai-agents › article
[ARTICLE · art-153332] src=unexcitedneurons.substack.com ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

Meta Muse gives you a pretty nice VM for free

A developer benchmarked the free Muse Secure VM that Meta provides to every Muse personal AI agent user and found it runs on an AMD EPYC 9D25 with 2 vCPUs (two SMT threads of a single physical core at ~1.5 GHz), 8GB RAM, and a 100GB home volume, with roughly 644 MB/s reads and 177 MB/s writes. Sustained testing over an hour showed no throttling, and OpenSSL SHA512 and AES-256 benchmarks scaled 1.98x and 2x across the two threads, indicating the vCPUs are SMT siblings rather than separate cores.

read2 min views55 publishedSep 30, 2026
Meta Muse gives you a pretty nice VM for free
Image: Unexcitedneurons (auto-discovered)

When Zuck launched Muse on September 8, most of the news focused on the agent. I got curious about the VM that it runs on. Every Muse agent gets its own Muse Secure VM, which is supposed to be a dedicated VM for each user, where your agent and data lives. This is available to every free user (not just paying users).

So I decided to ask what it is running on, and it turns out to be :

  • 2 vCPUs, clock at 1497.749 MHz, sharing 1 physical core
- AMD EPYC 9D25 126-Core Processor
- KVM (hypervisor flag set in /proc/cpuinfo)
  • 8GB of RAM
  • 100GB volume for home, plus 7.5GB mounted at /
  • About 644 MB/s read, 177 MB/s write
  • Ubuntu 24.04, Linux kernel 7.0

This is actually… quite great, especially for free. This would be a $10-20/month VPS at a typical provider. DigitalOcean would charge you $40/month for a similar machine.

Does Meta throttle the machine?

No.

Meta gives you full power as long as you need it. Unlike Amazon burstable instances (T2/T3/etc), I benchmarked the VM for over an hour, and never found any throttling.

I ran 2 workers for an hour minutes, logging throughput every second alongside CPU steal from /proc/stat. The first 30 seconds averaged 3,411,111 iterations/s; the last 30 seconds averaged 3,280,000. You get full power all the time, with no burst and no throttle.

Then I ran sysbench with 2 threads, and got only 1.44x the single-threaded score. That was ambiguous; it could be just constantly throttling me (instead of a burst throttle), or it could be 2 threads running via SMT on the same core.

So I ran OpenSSL SHA512 and AES-256 benchmarks, and got 1.98x and 2x respectively. This shows the 2 vCPUs are the 2 SMT threads of a single physical EPYC core, running at about 1.5 GHz.

For free, it’s definitely nice #

It’s just like having your own VPS box, with no shutdowns and no throttling. I saw no evidence that the VM was oversubscribed on the host. I’m sure it is, because I heavily doubt Meta would only put 126 VMs on a single server, but I never hit any limits.

── more in #ai-agents 4 stories · sorted by recency
── more on @meta 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/meta-muse-gives-you-…] indexed:0 read:2min 2026-09-30 · —