Alleged 14-trillion-param "Mythos-class" weights appear on Hugging Face, unverified
teortaxesTex · x · 2026-11-02
A repo claiming to contain internal weights of a 14-trillion-parameter "Mythos-class" model appeared on Hugging Face: 1,224 MoE experts, 7.72 TB of weights across 772 shards, hybrid GDN + MLA architecture, plus model configs, architecture code and training metadata. Nobody has verified the weights.
@teortaxesTex is skeptical: the architecture (hybrid GDN/MLA, 1,224 experts, NVFP4) doesn't quite match what would be expected of a next-gen Mythos-class run — "more effort than the last time at least" — and laments how quickly a copycat showed up. Unverified.
More from Models
- Quant-Arena benchmark: 5 of 9 LLM quant strategies lose money, none beat 17.1% baseline — jiqizhixin · 2026-11-02
- DeepSeek V4.1 Flash 1017 Scores 159 on Epoch ECI, Tied With GPT 5.5 — teortaxesTex · 2026-11-02
- Domingos: If OpenAI and Anthropic Had a Secret Sauce, Distillation Wouldn't Work — pmddomingos · 2026-11-02
- AI commentator laments frontier labs leak nothing, not even model configs — teortaxesTex · 2026-11-02
- First GPT-6.1 Terra test: perfectly rebuilds a scanned PDF form in 81 seconds — MikePFrank · 2026-11-02
- Polymarket puts 69% odds on GPT-6 "Astra" launching within two weeks — Polymarket · 2026-11-02