Z.AI raises about $5 billion, with most of it aimed at GLM-6.0 and a self-training system

Z.AI, the Beijing lab formerly known as Zhipu and the developer of the open-weight GLM family, completed a financing package worth about $5 billion, split between roughly $2 billion of new shares and $3 billion of convertible bonds, TechNode reported. The company says about 60% of net proceeds will fund research and development on the next generation of GLM foundation models and on what it calls its Fully Self-Training system, plus deployment and upgrades of large-scale training and production-inference compute. On its first earnings call as a listed company on August 31, founder Tang Jie named GLM-6.0 as the next base model and described the self-training approach as covering pretraining, midtraining and post-training, with the system judging when to stop or correct itself.

The self-training claim deserves caution. Recursive self-improvement is asserted far more often than it is demonstrated, and no external party has yet evaluated whether Z.AI's loop produces gains beyond conventional data curation and reinforcement learning on model-generated trajectories. A paper posted to arXiv on Tuesday, Generalized Agent Iteration, makes the same point academically, observing that self-improvement is claimed at many scales without a shared formal framework to say what is being claimed.

What is measurable is how much the open-weight tier has improved. A new arXiv paper introduces ZGCM-1, a fully open 7-billion-parameter dense model trained from scratch, released with weights from the pretraining, midtraining and post-training stages, intermediate checkpoints, training code, per-stage data recipes and training logs. The authors report 75.0% on AIME 2026, 97.1% on MATH-500 and 70.4% on HMMT 2025, and claim the top average across 14 reasoning benchmarks in the 7-to-8-billion-parameter class, with competitive results against models orders of magnitude larger including Qwen3-235B-A22B and GLM-5.1. The design pairs interleaved gated sliding-window and full attention with an FP8 Muon optimizer and a context curriculum that scales from 16,000 to 256,000 tokens. These are author-reported numbers on public benchmarks with known contamination risk, so independent reruns matter.

Taken together, the two items describe the same market. Capital is flowing into Chinese frontier training at multibillion-dollar scale while the open small-model tier is absorbing capability that previously required a large closed model and a metered API.