GLM 5.3 Flash: Open-Weights Frontier-Adjacent AI

Model Architecture and Performance

GLM 5.3 Flash is a mixture-of-experts model featuring 320 billion total parameters with 18 billion active parameters per token. It utilizes a hybrid attention mechanism, manifold constrained hyperconnections, and an index pool technique to maintain performance across a 1 million token context window. In independent testing on KingBench, the model achieved a score of 63 out of 80 (78.75%), placing it competitively against models like Opus 4.8 and DeepSeek V4 Pro. While the model excels at reasoning, math, and agentic coding tasks—such as generating datasets and performing local fine-tuning via Apple's MLX framework—it showed slight performance variance in visual one-shot generation tasks compared to its initial stealth preview as "Ox Alpha."

Local Inference and Hardware

The model's architecture is optimized for local deployment, as the 18 billion active parameters ensure high inference speeds while the 320 billion total parameters dictate memory requirements. At 4-bit quantization, the model requires approximately 180 GB of VRAM, though upcoming 2-bit and 3-bit quantizations from Unsloth are expected to reduce this footprint toward the 100 GB range. This makes the model a primary candidate for new high-memory hardware, such as Apple's M6 Max and M5 Ultra Mac Studio systems, or specialized AI boxes like the Xiaomi AI Cube, which provide the unified memory and bandwidth necessary to run frontier-adjacent models on-device.

Infrastructure and Economics

Z AI served the entire stealth preview of 44 trillion tokens using a custom inference stack built on SGLang, running exclusively on Chinese AI chips. This infrastructure choice allowed for a 3x improvement in serving performance and supports the current API pricing of $0.15 per million input tokens and $0.50 per million output tokens. The model is available under an MIT license, facilitating broad adoption for local agentic workflows and private fine-tuning.