Who Is Quietly Building China’s Mythos‑Level AI? Musk Says 9 Months, Tang Says It’s Not That Fast

Mythos‑level models have been banned, leaving China to rely on domestic development. Elon Musk claimed a nine‑month timeline to catch up, but ZhiPu chief scientist Tang Jie replied that it won’t be that quick.

The discussion shifts to whether advancing solely through pre‑training is sufficient. OpenAI’s jump from GPT‑4 to o1 and Anthropic’s Constitutional AI both rely heavily on post‑training, which can unlock a larger portion of a model’s potential beyond the base architecture.

Mind Lab, a team under Mindverse, is currently the only external organization that has completed full‑process post‑training for the GLM‑5.1/5.2 series. Their work on the Macaron‑V1‑Preview model already surpassed the base GLM‑5.1 scores, indicating substantial room for improvement through post‑training.

Benchmark results show that post‑training on GLM‑5.1 raised PinchBench from 76.6 to 92.5 (a 15.9‑point, ~20.8% relative increase) and Terminal‑Bench 2.0 from 63.5 to 67.4 (a 3.9‑point gain). These figures demonstrate that the GLM base still has untapped capacity.

With the release of GLM‑5.2, which introduces the IndexCache architecture, Mind Lab quickly adapted their post‑training pipeline and open‑sourced support for the new features, including adaptations for Dynamic Sparse Attention (DSA) and Multi‑Token Prediction (MTP) used in models over 700 B parameters.

The team’s rapid response—often within days of a new base model release—highlights a key advantage: post‑training cycles are measured in weeks, far shorter than the months‑long pre‑training cycles. This speed could allow external teams like Mind Lab to close the gap to Mythos‑level performance faster than expected.

However, the post‑training path faces high barriers: deep understanding of the base architecture, high‑quality training data construction, and robust engineering infrastructure. Mind Lab has addressed these by releasing a Megatron‑compatible training framework for GLM‑5.1/5.2 and demonstrating end‑to‑end post‑training pipelines.

In summary, while the nine‑month claim remains uncertain, Mind Lab is presently the only visible external team with a proven post‑training track record on the latest GLM bases, positioning it as a critical player in China’s quest for Mythos‑level AI.