GLM-5.3 Analysis Examines Chinese AI Lab Progress
Nathan Lambert examines why GLM-5.3 performance should no longer surprise observers of Chinese labs.
Nathan Lambert posted notes on GLM-5.3 from Chinese labs. He argues the model's strength reflects real technical gains rather than distillation from Western systems. Replies discuss inference compute gaps between US and Chinese labs. Benjamin Warner called for open source developers to match the safety testing expected from OpenAI and Anthropic. Other posts note data purchases by Chinese firms and post-training refinements on the GLM-5.2 base.
GLM 5.3 notes and why we should stop being so surprised about these very strong Chinese models (most of this is talking myself through some of my denial -- yes, these models are the real deal).
Reactions from ranked influencers
13 postsWith prodding from @xeophon I added a crucial detail. Data industry go brrr in China. Our interconnects group chat has on many occasions been discussing the data industry in China recently, a huge change from when we visited in April.
GLM 5.3 notes and why we should stop being so surprised about these very strong Chinese models (most of this is talking myself through some of my denial -- yes, these models are the real deal).
2+ years after me and @zephyr_z9 started to talk about this. Well. Seeing how far they’ve come on purely technical progress and distillation, I guess it wouldn’t be interesting otherwise. I expected a much greater gap due to data inferiority (Chinese web is also impoverished)
With prodding from @xeophon I added a crucial detail. Data industry go brrr in China. Our interconnects group chat has on many occasions been discussing the data industry in China recently, a huge change from when we visited in April.…
@teortaxesTex @zephyr_z9 Well it really wasn’t a thing up until a few months ago. Like that’s the real distillation happening, not some stolen CoTs
@xeophon @teortaxesTex Startups started popping up 12-15 months ago Tencent, Alibaba, BD etc have bought data from them for a long time Zhipu, kimi et al. started buying data from Q4 25/Q1 26 Look into EnableExpert (Mercor clone) launched in April 2026
@zephyr_z9 @teortaxesTex it’s about CN labs and companies buying from US vendors, that thing is rather recent-ish April 2026 is very recent as well, Unipat also just started end of 2025/this year (but grew very very quickly)
Now that OSS models are pretty good, we need OSS labs to do the same safety testing we expect OpenAI and Anthropic to do.
GLM 5.3 notes and why we should stop being so surprised about these very strong Chinese models (most of this is talking myself through some of my denial -- yes, these models are the real deal).