Independent Security Researcher

Z.ai's GLM-5.3-Flash is the first natively multimodal GLM-5 model: 320B total / 18B active, hybrid sparse + linear attention, $0.15 per million input tokens, and MIT weights on day one. It beats GLM-5.2 on every published benchmark at a tenth of the price — and it's the anonymous "Ox Alpha" that took over OpenRouter.

Z.ai shipped a major capability jump without changing the base model at all — and then held the open weights back because the same training made it dangerously good at exploiting vulnerabilities. Benchmarks (vendor and independent), the 2,436 real bugs it found, every price and endpoint.

China passed the US in global AI model downloads and now sets the world's price floor for tokens. The plain-English case — open weights, structural cost efficiency, domestic silicon, ecosystem density — plus the honest caveats.

DeepSeek's V4 model lineup, the open-source DeepSeek Harness agent framework, and every DeepSeek API off-peak discount window and price — with sources.