Rumor: alleged Anthropic 'Mythos-class' 14T-param model leak appears on Hugging Face, unverified
jmorant555 · reddit · 2026-11-02
A Reddit post claims a 'Mythos-class' model has been leaked on Hugging Face (repo burner22/numbat), allegedly an internal Anthropic training checkpoint. Per the attached screenshot, the claimed specs are 14 trillion parameters, hybrid GDN + MLA architecture, 1,224 MoE experts, 7.72TB of weights across 772 shards, NVFP4 quantization, plus config files, architecture code, and training metadata. Unverified — the post itself hedges with 'allegedly' and 'big if true,' and it could well be an elaborate fabrication.
More from Models
- Rumored 'gpt-6.2-astra' briefly surfaced in Codex CLI, unverified — Angaisb_ · 2026-11-02
- Developers chafe at Claude Code's weekly limits: 75% used, 3 days to reset — adamamcbride · 2026-11-02
- Alleged 14-trillion-param "Mythos-class" weights appear on Hugging Face, unverified — teortaxesTex · 2026-11-02
- Fable 5.5 with Opus-Next routing generates realistic 3js pelican biking animation in one shot — inductionheads · 2026-11-02
- LiveOverflow jokes nobody asked him to post-train open models on exploit data — kingslyj · 2026-11-02
- Researcher mocks Apple's 'smarter' Siri as impossibly brain-dead in 2026 — MattGarciaEth · 2026-11-02