Z.ai's GLM-5.3-Flash Matches Top Models at 7.5x Lower Cost, Runs Entirely on Chinese Chips

Z.ai releases a cheaper, faster GLM-5.3 variant that runs on non-Nvidia hardware

Z.ai has released GLM-5.3-Flash, a 320-billion-parameter model with only 18 billion active parameters, a one-million-token context window, and an MIT license. According to Z.ai, it is the first natively multimodal model in the GLM-5 series. Weights are available on Hugging Face.

On Artificial Analysis's Intelligence Index, GLM-5.3-Flash scores 57 points at maximum reasoning effort, just three points behind the larger GLM-5.3 (60) and roughly level with GPT-5.6 Terra and Muse Spark 1.2. The gap in price is far larger than the gap in score: cost per task on the index runs $0.09 for GLM-5.3-Flash versus $0.68 for GLM-5.3, a 7.5x reduction. Artificial Analysis places the model on the Pareto frontier of intelligence versus cost.

On Z.ai's API, GLM-5.3-Flash is priced at $0.15 per million input tokens and $0.50 per million output tokens, about ten percent of GLM-5.3's rate. On agentic benchmarks the smaller model holds up well: on GDPval-AA v2 it posts an Elo score of roughly 1770, matching both GLM-5.3 and Grok 4.6, and trailing only Claude Opus 5. Artificial Analysis notes a tradeoff, however — the model is less token-efficient, with about 90 percent of its output tokens spent on reasoning rather than final answers.

Tested anonymously, served on Chinese chips

Before its public launch, Z.ai reportedly tested the model anonymously under the codename "ox-alpha" on OpenCode and OpenRouter, where it became the most-used model of the week. According to Z.ai, all of that traffic ran on Chinese AI chips rather than Nvidia hardware. SemiAnalysis reports the deployment served 100 trillion tokens per day, a capacity level previously associated only with frontier labs running on Nvidia infrastructure.

Z.ai claims its hardware efficiency and cost per token on this setup are on par with common Nvidia GPUs. SemiAnalysis frames this as another data point testing Nvidia's "CUDA moat" — the software layer that has tied nearly two decades of AI framework development to Nvidia hardware. Moving workloads to other chips typically requires reprogramming compute operations, adjusting memory access patterns, and resolving new bottlenecks from scratch.

To make the switch work, Z.ai says it built custom serving software on top of SGLang, splitting inference into independently scaling stages. The company claims this approach tripled throughput compared to its first attempt on the same hardware, with a GLM-5.3-based agent assisting in the optimization work itself.

What this means

GLM-5.3-Flash is a concrete data point in the broader push by Chinese AI labs to compress the cost of near-frontier intelligence. A model landing within three points of its larger sibling on a standard benchmark while costing 7.5 times less per task is a meaningful efficiency gain, not a marginal one, and it adds pressure on Western providers whose flagship pricing has not moved nearly as fast.

The more consequential claim, if verified independently, is the non-Nvidia serving infrastructure. Sustaining 100 trillion tokens a day on Chinese chips with custom SGLang-based software suggests the CUDA moat is narrower than assumed for at least some large-scale inference workloads. That doesn't mean Nvidia's training-side dominance is at risk, but it does indicate that serious alternatives for inference at scale are closer to production-ready than commonly believed — a development worth watching regardless of how the benchmark numbers hold up under further scrutiny.