GLM-5.3 built its own inference stack, on 100K Chinese chips
Recursive self-improvement, or good tooling? Either way it tripled throughput in two weeks — why GLM-5.3-Flash is $0.075/M. Plus CrowdSec's code leaks.
Copy markdown
GLM-5.3 wrote the code that serves it
Z.ai says its own GLM-5.3 model did most of the engineering — scheduling, kernels, optimization — on the inference stack now running GLM-5.3-Flash in production. It went from first successful run to production-ready in under two weeks, with end-to-end throughput tripling over the baseline through a 'dense feedback' loop between model and infra.
100,000 Chinese chips, NVIDIA-class economics
The cluster runs on 100,000+ domestic Chinese accelerators — the first deployment at that scale — with per-token cost and hardware efficiency Z.ai claims are 'comparable to mainstream NVIDIA GPUs.' It served 62 trillion tokens in the six days after launch. The takeaway for builders: serious capacity without an NVIDIA dependency.
Why you care: a 1M-context model at $0.075/M
GLM-5.3-Flash is 320B total / 18B active, natively multimodal, 1M-token context, and ships MIT open weights. It runs about $0.075 per 1M input tokens on OpenRouter and is already the platform's most-used model. This infra work is why that price and capacity hold up under load — cheap long-context calls you can lean on for agents and batch jobs.
Read the 'recursive self-improvement' label skeptically
This is a first-party blog post, and 'the model built its own infra' is Z.ai's framing: GLM-5.3 assisted human engineers, it didn't act alone, and the RSI branding is doing marketing work. The throughput numbers are concrete and useful; the AGI-adjacent narrative is unverified.
Elsewhere: CrowdSec's private source code leaks
CrowdSec confirmed ~300 private repos — its SaaS console, AWS routines, connectors — were exposed after a backdoored dependency (tied to the Tanstack supply-chain compromise) stole a read-scoped API key back in May; it surfaced Sept 16. No customer data or the open-source Security Engine was hit, and tokens are rotated. The real lesson is the dependency vector, not the leak itself.