GLM-5.2 open-sourced by MIT, the first time globally that a practical intelligent threshold has been crossed, and an open-source model that can be infinitely modified.

Some startups are already getting started. The most focused discussion is interesting: not on coding, but on moving toward long-term agentic RL algorithms outside of coding. Healthcare, legal, finance, manufacturing... countless vertical fields with long-term tasks, previously stuck in papers due to insufficient base model capabilities + inability to freely modify. GLM-5.2 lowers the barrier to an open-source playable level.
If early this year OpenClaw accelerated the first consensus on agentic (solving "can it move"), then GLM-5.2 is the second acceleration (solving "can it continuously operate in complex environments"). The difference is—OpenClaw is a lab toy, GLM-5.2 is an industrial-grade base.
MIT’s three statements: freely modify, freely sell, and the only obligation is to retain the copyright notice. This is the premise for the "mass after-training movement" to succeed. Zero intellectual property risks, commercially usable, embeddable in business, no need to disclose modifications.
Especially with recent events—the flagship model of Anthropic was required to be taken offline globally due to US export controls. "Domestic models + domestic computing power" shift from alternative to mandatory. On the first day of GLM-5.2 release, all mainstream domestic computing platforms were adapted, and the stock price roughly doubled in a week.
The biggest benefit isn’t from GPUs themselves (open source + adaptation by 8 domestic platforms → inference chip deflation). The real inflation pressure is in four layers (these four leading stocks are very clear, with AI mining some niche stocks that haven't surged significantly, higher risk):
🥇 HBM—most critical globally
Long-term agent multi-hop reasoning causes exponential increase in bandwidth consumption. The three major manufacturers’ capacities are sold out, with a 50%-60% gap.
Wanrun Technology (002654): partnered with Yangtze Memory to complete 12-layer HBM3E package adaptation for Nvidia H200, securing the first batch of 300,000 units. The actual controller is under Hubei State-owned Assets Supervision and Administration Commission, same as Yangtze Memory. UBS and Morgan Stanley bought in during Q1.
🥈 Optical chips/InP—CPO inflation amplifier
InP substrate shortage over 70%.
Yunnan Geology (002428): subsidiary XinYao Semiconductor is a leading InP substrate manufacturer, certified by Nvidia. The only domestic 6-inch InP substrate in mass production, with a yield of over 70%. Huawei Hubble invested. Ge and indium phosphide are both scarce.
🥉 ABF substrates—certainty of shovels
Ajinomoto ABF film price increased by 30%.
Hongchang Electronics (603002): GBF multilayer film samples sent to TSMC and Changdian Technology.
🥄 CCL M9—material layer inflation
M9 price is 10 times that of ordinary FR4.
Nanya New Materials (688519): M6-M8 supplied in bulk, M9 in progress. Small scale, high flexibility.
GLM-5.2 has pulled the trigger. The bullets are still flying.