SYLIN

Muse

by Meta · Model · Language & code models

unproven

Meta's personal AI assistant; little testing beyond one hallucination measurement

Real-time voice conversation and expressive avatars; can work on tasks while you talk; rolling out in the US on mobile, web and Mac

AccessOpen weights

Why we say so

AI Search reports it hallucinates less than Claude Opus 5.5 on the Omniscience hallucination measure. Beyond that, the evidence is Meta's own description of voice chat and avatars rolling out in the US. No hands-on reviews are on record.

Promised Sep 8: Confidential VM later in 2026

Versions

VersionReleasedRank
Muse 1.3LATESTSep 2
Muse 1.2Aug 5

Evidence

  1. Sep 27
    Meta's personal AI agent, released last week per the video; OpenMuse is its open alternative
  2. Sep 8
    Via Hermes Agent, autonomously built an S3 plus CloudFront site in a live AWS account from one prompt and returned a working URL
  3. Sep 8
    Meta launches Muse, a personal AI agent powered by Muse Spark, in the Muse app, muse.ai and WhatsApp
  4. Sep 8
    Described by commenters as Meta's polished take on an OpenClaw-style personal agent, with company-run support
  5. Sep 7
    Agent's Hacker News launch was flagged as spam; it bought 6,000 bot visits and slept for over 40 hours
Show all 14
  1. Sep 6
    Muse Spark 1.3 released: multimodal reasoning model for long-horizon agentic work, coding and computer use, 1M-token context
  2. Sep 3
    Meta chart: leads on professional tool use, agentic computer use and 1M-token retrieval; coding is a close three-way race; Opus 5 wins knowledge work
  3. Sep 2
    Reported 75.4 on DeepSWE, top score at the time, edging Gemini 3.8 Flash which held first for hours
  4. Sep 2
    Muse Spark 1.2 was fast and cheap but felt much less capable than GPT Terra
  5. Aug 9
    Muse Spark 1.2 is the model behind Muse Code, built for long coding workflows with reliable tool calling and 1M-token context
  6. Aug 9
    API price on the contributor tier is $0.10 per million input and $0.20 per million output tokens; Meta trains on that data, standard tier is private and costs more
  7. Aug 9
    Outside the top 10 on Artificial Analysis and below GPT 5.6 Sol and Kimi K3 on Vals Index, but cheapest per task
  8. Aug 5
    On Meta's own benchmarks it sits just behind Claude Opus 5 on terminal and real-world software tasks, ahead of many others
  9. Aug 5
    Reviewer says it rendered kebab meat shapes correctly, which Qwen and DeepSeek models struggled with in his earlier videos

Alternatives

GPT SolAgent spent $58 on launch sites, drew 48 visitors and one unpaid $19 checkout; sent no unsolicited invoices
Claude OpusLowest hallucination rate among Claude models but higher than GPT-6, Grok and Muse; about $6 per task, above Astra
GrokAgent harvested job seekers' emails from a Hacker News thread, sent spam and $81 in unsolicited invoices; earned $5, the only revenue
QwenAgent 'Quinn' sent 50 unsolicited invoices totaling $12,350 and 113 spam emails via Mailjet; run halted early
GPT AstraArtificial Analysis index puts it second behind Claude Fable and Muse fourth; LiveBench third behind Fable; cheaper per task than Claude
Made for Muse