The stealth model 'ox-alpha' that topped OpenRouter's popularity charts turned out to be Z.ai's GLM-5.3 Flash, a 320B-parameter (18B active) hybrid model built for ultra-low-cost inference. Released under MIT license on Hugging Face, it scores 57 on the Artificial Analysis Intelligence Index, roughly matching GPT-5.6 Terra and Gemini 3.7 Flash, and performs even better on agentic benchmarks, approaching Claude Opus 4.8 at a fraction of the cost ($0.075/$0.25 per million input/output tokens). Notably, Z.ai served the model on Chinese AI chips rather than Nvidia GPUs, using a custom SGLang-based inference engine and a hybrid linear/sparse attention architecture that cuts compute by 3x and KV cache size by 4.4x versus the full GLM 5.3 model, achieving performance comparable to mainstream Nvidia hardware.
Table of contents
Benchmarks: It’s good, but not FableCheap inference on Chinese chipsLinear and sparse attentionQuestions this post answers
What model was hidden behind the ox-alpha stealth model on OpenRouter?
Ox-alpha was Z.ai's GLM-5.3 Flash, a 320 billion-parameter hybrid model with 18 billion active parameters, unmasked and released with weights on Hugging Face under the MIT license. It quickly became the most popular model on OpenRouter before its identity was revealed. Developers tracking which stealth models turn into real releases can follow model launches like this on daily.dev.
How does GLM-5.3 Flash pricing and performance compare to Claude Opus 4.8?
GLM-5.3 Flash can match Claude Opus 4.8 on benchmarks that matter for AI agents while costing roughly a tenth of the price, priced at $0.075 per million input tokens and $0.25 per million output tokens on OpenRouter (with a 50% discount applied). It scores 57 on the Artificial Analysis Intelligence Index, in line with GPT-5.6 Terra and Gemini 3.7 Flash. Teams weighing model cost against agentic performance can compare launches like this one on daily.dev.
How did Z.ai achieve low-cost inference for GLM-5.3 Flash on Chinese AI chips?
Z.ai built a custom inference engine based on SGLang and used its flagship GLM-5.3 model to power an infrastructure agent that helped optimize the serving system itself, achieving a 3x improvement in end-to-end serving performance over its initial baseline on the same Chinese chip hardware. The model also uses a hybrid linear and sparse attention architecture that cuts compute by 3x and KV cache size by 4.4x compared to the full GLM 5.3 model. Engineers evaluating non-Nvidia inference stacks can keep up with efficiency breakthroughs like this on daily.dev.