Leaker estimates GPT-6 Astra at 4.2T params, 112 layers with 50% layer looping
scaling01 · x · 2026-09-28
- Leaker scaling01 shares his latest central estimate for GPT-6 Astra: 4.2T total params, 120B active, 112 layers, with 50% of layers looped once (a looped language model).
- He claims evidence suggests GPT-6 Sol, Luna, and Opus 5.5 are not looped models, and withheld his full methodology for further analysis.
- He adds that if the model is much smaller (3T), it might loop 2-3 times instead, given otherwise implausible 2% active ratios.
Unverified speculation, for reference only.
Related event: Leaker estimates GPT-6 'Astra' as 4.2T-parameter looped language model(5 posts)→
More from Models
- Apple dictation users report recent quality regression with random word errors — whatsallthiss · 2026-09-28
- TypeSafe AI's Jev returns confidence-scored decisions, not text: a16z sits down with Diogo Almeida — a16z · 2026-09-28
- Zvi jokes Opus 5.5 totally missed an obscure cultural reference while editing his blog — TheZvi · 2026-09-28
- Meta recaps 6 months of launches: 10 Muse releases plus Meta Model API — AIatMeta · 2026-09-28
- Vercel CEO: open models now drive 80% of Vercel's token traffic as OpenAI, Gemini, Anthropic lose ground — RihardJarc · 2026-09-28
- Real-time robot control showdown: GPT-6 Astra vs Opus 5.5 vs Grok 4.7 vs MolmoAct2 — DJiafei · 2026-09-28