Developers probing a free mystery model across AI services were testing an early build of a new Z.ai system that the company later released with better performance and stability. Z.ai said the model, known publicly as Ox Alpha, was GLM-5.3-Flash, the first multimodal model in its GLM-5 series, released under the MIT license with a 1M token context window and 320B total parameters, 18B active at inference.
OpenRouter said Ox Alpha became the biggest model ever on its marketplace, processing more than 20T tokens in 6 days. Z.ai put GLM-5.3-Flash on a two-week 50% discount through its API, lowering prices to $0.075 per 1M input tokens and $0.25 per 1M output tokens, and said the preview ran entirely on Chinese AI chips. In later testing, GLM-5.3-Flash ranked second on ErdosBench behind GPT-5.6 Sol xhigh.
Key sources
- SOURCE@iscienceluvr“Remember people saying the model was Google, Meta, even SSI”x.com
- SUPPORT@theturingpost“users gave it ~8 million sessions and 24 trillion tokens of real-world testing”x.com
- SUPPORT@prz_chojecki“GLM-5.3 Flash finished 2nd on our ErdosBench right after GPT-5.6 Sol xhigh.”x.com
- SUPPORT@theo“It’s so good that I had to move other things around.”x.com
- SUPPORT@baseten“All served from the US only, starting on day 0, with ZDR by default.”x.com