Z.ai releases GLM-5.3-Flash, a 320B open-weight model that ran on Chinese AI chips

By NewSunSEO | September 23, 2026

Z.ai has released GLM-5.3-Flash, a 320 billion parameter model that scores 57 on the Artificial Analysis Intelligence Index for roughly $0.09 per task, about 7.5 times cheaper than the larger GLM-5.3, and it ran its full anonymous preview on Chinese AI chips rather than Nvidia hardware.

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series, according to Z.ai. It has 320 billion total parameters, of which only 18 billion are active at any time. It ships under an MIT license, supports a context window of one million tokens, and has its weights published on Hugging Face.

How does GLM-5.3-Flash compare on benchmarks?

Artificial Analysis measured GLM-5.3-Flash at 57 points on the Intelligence Index at maximum reasoning effort. That puts it three points behind the larger GLM-5.3, which scores 60, and level with GPT-5.6 Terra and Muse Spark 1.2. On agentic work, the model reaches an Elo score of about 1770 on GDPval-AA v2, matching GLM-5.3 and Grok 4.6, with only Claude Opus 5 ahead of it.

The price difference is wider than the quality gap. Cost per task on the index runs $0.09, against $0.68 for GLM-5.3, roughly 7.5 times cheaper. On Z.ai’s own API, input tokens cost $0.15 per million and output tokens cost $0.50 per million, a little over 10% of the price of GLM-5.3. Artificial Analysis places the model on the Pareto frontier of intelligence and cost.

What are the model’s weaknesses?

GLM-5.3-Flash is less token-efficient than its larger sibling. Artificial Analysis found that roughly 90% of the output tokens the model produces go to reasoning, meaning it spends more of its output budget thinking through a task than some competing models do.

Why did the model run on Chinese AI chips instead of Nvidia?

Before launch, Z.ai tested the model anonymously under the name “ox-alpha” on OpenCode and OpenRouter, where it became the most popular model of the week. All of that traffic ran on Chinese AI chips, according to Z.ai. SemiAnalysis reports the test traffic reached 100 trillion tokens per day, a level of serving capacity that until now was thought possible only for frontier labs.

Z.ai claims its hardware efficiency and cost per token land on par with common Nvidia GPUs. SemiAnalysis frames the result as another test of Nvidia’s CUDA moat, the nearly 20-year-old programming layer between AI software and graphics cards that most frameworks are tuned for. Switching away from CUDA means rewriting compute operations, adjusting memory access, and hunting down bottlenecks by hand.

To make the non-Nvidia stack work, Z.ai built its own serving software on top of SGLang and broke processing into stages that scale independently. The team says this approach tripled throughput over its first attempt on the same hardware. An agent based on GLM-5.3 helped with the optimization.

Where does GLM-5.3-Flash sit in the price pressure on Western providers?

The $0.09 per task figure adds GLM-5.3-Flash to a growing list of Chinese models that have recently applied heavy cost pressure on Western providers. The model delivers benchmark scores within three points of a frontier-class system while charging a fraction of the price, which Artificial Analysis reads as the new competitive floor for cost-versus-intelligence tradeoffs.

FAQ

What is GLM-5.3-Flash?

GLM-5.3-Flash is a model released by Z.ai with 320 billion total parameters and 18 billion active parameters. It has a one million token context window, ships under an MIT license, and is the first natively multimodal model in the GLM-5 series. Its weights are available on Hugging Face.

How much does GLM-5.3-Flash cost?

On Artificial Analysis’s Intelligence Index, GLM-5.3-Flash costs about $0.09 per task, roughly 7.5 times cheaper than GLM-5.3. On Z.ai’s API, it costs $0.15 per million input tokens and $0.50 per million output tokens, a little over 10% of the price of GLM-5.3.

Did GLM-5.3-Flash really run on Chinese AI chips?

Yes. Z.ai says the model’s anonymous preview under the name “ox-alpha” on OpenCode and OpenRouter ran entirely on Chinese AI chips. SemiAnalysis reports the test traffic reached 100 trillion tokens per day, and Z.ai built its own serving software on top of SGLang to triple throughput on that hardware.

Related coverage

• Cisco Antares: Open-Weight AI Models for Vulnerability Localization
• Qwen3.6-27B: Open-Source 27B Model Beats 397B MoE on Coding Benchmarks
• Meituan’s LongCat-2.0 Was Trained Entirely on Chinese Chips, the Company Says