Speculation: OpenAI o1 may be GLM-based; mysterious lab's efficiency praised

On August 21, user teortaxesTex posted a series of threads offering personal speculation about the underlying architecture of OpenAI's o1 and the capabilities of a certain "mystery lab" — all unverified community opinions with no official confirmation.

Confirmed

• teortaxesTex hands-on tested O1 Alpha in the DSH project, claiming it can handle not only routine code but also complex tasks such as merging projects and adding ZBrush features, with a pragmatic chain-of-thought process and performance approaching top-tier models.
• Based on that test, the author speculates that o1 (ox alpha) may run on a GLM architecture version 5.x with vision capabilities, rather than being simply a 1T-parameter or other-scale model.
• The author praises a certain mystery lab for its extremely high compute-to-results efficiency, able to quickly ship frontier models at the Mythos-Preview and Mythos 2 level ahead of the author's expected timeline of late 2026, and predicts the lab won't hit bottlenecks before reaching — or even surpassing — the US "frontier" level.
• The author adds that the lab's current results are actually catching up to US checkpoints from a few months ago; internally it has strong models capable of solving math problems, but these builds rely on US models from several months earlier.

Not Yet Confirmed

• Whether o1 is truly based on the GLM architecture is purely the author's personal speculation, unsupported by any official or first-hand evidence.
• The identity of the "mystery lab" and its model roadmap likewise remain unverified.

Why It Matters

This series of speculations reflects the community's strong curiosity about o1's architectural origins and training路线. If the GLM architecture guess holds up, it would significantly reshape perceptions of where frontier-model technology comes from; meanwhile, the "compute conversion efficiency" lens offers a novel community framework for evaluating labs' competitiveness. Readers should note, however, that everything above is the speculative opinion of a single author and should not be cited as fact.

2026-08-21 ~ 2026-08-21 · 5 related posts

• [source] Speculation suggests OpenAI o1 uses GLM architecture, rivaling Claude — teortaxesTex · 2026-08-21
• [source] Tests suggest O1 Alpha is a GLM-based model with advanced coding capabilities — teortaxesTex · 2026-08-21
• Lab praised for high compute efficiency and rapid model delivery — teortaxesTex · 2026-08-21
• Mysterious Lab Expected to Rapidly Deliver Frontier-Level SWE Models — teortaxesTex · 2026-08-21
• [source] Internal Models Strong Yet Rely on Older US Base Models — teortaxesTex · 2026-08-21