Leak: OpenAI Lagged in Pre-training, Developing 10T Param Model 'Astra'
iruletheworldmo · x · 2026-08-10
According to leaks, OpenAI realized it was lagging in pre-training, prompting resource cuts to projects like Sora to concentrate compute. They are now advancing a massive pre-train codenamed Doug.
The next-generation model based on this pre-train is codenamed Astra, with an estimated 10T parameters. Researchers are reportedly much more excited about Astra than current models. The previously known Sol model was 5T parameters.
Related event: OpenAI's 10T-Parameter Model Astra Reportedly Launching in August(6 posts)→
More from Models
- Muse Glimmer Open-Sourced: A 30B Dense Model for Agentic Use Cases — JesseDodge · 2026-08-10
- Running 30B Model on Single RTX 5060 Ti with 131k Context Window — BazzyIm · 2026-08-10
- Topology Test: Grok Imagine 2.0 Beats ChatGPT by Correctly Understanding Genus — luismbat · 2026-08-10
- Testing LLM Reasoning: A Solar Eclipse Probability Math Problem — Afinetheorem · 2026-08-10
- OpenAI Locks Down Astra After Model Raises First-Ever Critical Cyber Capability Fears — sksarkpoes3 · 2026-08-10
- Opinion: Does Google Need a Top-Tier Coding Model? Skipping to Gemini 4 — haider1 · 2026-08-10