Z.ai unveils GLM-5.3-Flash: a high-performance multimodal model with enhanced capabilities

Z.ai has launched GLM-5.3-Flash, a new model in its GLM-5 series that is natively multimodal and designed for efficient coding tasks. The model is a mixture-of-experts (MoE) architecture with a total of 320 billion parameters and 18 billion active parameters per token, featuring a context window of 1,048,576 tokens. It supports image and video inputs and is released under an MIT license, with its weights available on Hugging Face. According to Z.ai, GLM-5.3-Flash outperforms its predecessor, GLM-5.2, across various benchmarks and real-world workloads at roughly one-tenth the cost. It also scores within half a point of Claude Opus 4.8 on its internal coding benchmark. The model initially operated under the anonymous name 'Ox Alpha' on platforms like OpenCode and OpenRouter, running on domestically produced Chinese AI chips. Z.ai reports that the Ox Alpha preview utilized a custom SGLang-based engine, which disaggregates encoding, prefill, and decoding processes, resulting in a 3× improvement in end-to-end serving performance across tens of thousands of accelerators. GLM-5.3-Flash is based on a newly trained base model using a 30T-token multimodal corpus. Three key changes are highlighted: the model's performance on benchmarks, its pricing structure, and its multimodal capabilities. The model's API pricing is set at $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens. On the discounted tier, Z.ai reports a score of 57 on the Artificial Analysis Intelligence Index v4.1.1 at a rate of $0.045 per task. The model is available across all GLM Coding Plan tiers—Lite ($18/month), Pro ($80), and Max ($168)—with three times the usable quota of GLM-5.3. Its multimodal features are accessible through ZCode via Browser Use and Computer Use. Local serving is supported on several frameworks, including SGLang, vLLM, TokenSpeed, and KTransformers. Vision capabilities, however, are noted as a weaker area, with the model trailing Gemini 3.7 Flash on BabyVision and MVbench. Artificial Analysis independently scores the model 57 on the Intelligence Index, with a throughput of 48.7 output tokens per second and a 1.52-second time-to-first-token (TTFT) on Z.ai’s API. While the model demonstrates strong value for its cost, it is noted as being slow and verbose. The model's weights and blog are available for further details, and Z.ai encourages engagement through its social media channels and community platforms.