Zhipu Open-Sources Native Multimodal Large Model GLM-5.3-Flash: 320B Parameters and Full-Stack Support from Domestic Chips
Zhipu officially releases and open-sources the world's first native multimodal large model GLM-5.3-Flash. The model features a total parameter count of 3200亿, with only 18B activated parameters, delivering comprehensive performance comparable to Claude Opus 4.8. Leveraging a hybrid sparse and linear attention architecture along with multiple memory optimization techniques, the model achieves significant breakthroughs in long-context processing and service overhead. Inference services are fully supported by domestic AI chip clusters, drastically reducing costs. It is now globally open-source and offers enterprise-level API access.
Zhipu officially open-sources GLM-5.3-Flash (320B-A18B). As the first native multimodal large model in the GLM-5 series, its total parameter count reaches 3200亿. It scores 57分 in the Artificial Analysis Intelligence Index evaluation, with overall capabilities comprehensively surpassing previous generations, and performs comparably to Claude Opus 4.8 in self-developed programming benchmarks.
The model achieves significant efficiency optimization in architectural design. By introducing a hybrid paradigm of sparse and linear attention, combined with the innovatively developed IndexPool weighted pooling technology, it effectively resolves computational bottlenecks in long-context scenarios. Comparative experiments show that its single-Token computation volume and KV cache size are reduced by approximately 3倍 and 4.4倍 respectively compared to the previous version, placing its comprehensive computational efficiency at the forefront of the industry.
To adapt to domestic hardware environments, Zhipu has built a dedicated inference engine deeply customized based on SGLang, assisted by an Infra Agent driven by GLM-5.3 to complete CUDA operator development and performance tuning. The system integrates aggressive memory optimization strategies such as W8A8 weight quantization, mixed-precision KV caching, and Layer Split hierarchical offloading, successfully overcoming the challenge of limited memory bandwidth in domestic chips.
At the cluster deployment level, the team adopts a production-grade Encode-Prefill-Decode decoupled architecture, decoupling multimodal encoding, prompt prefilling, and token-by-token decoding into independent work pools. Real-world test data shows that this solution improves end-to-end service performance by 3倍. The single-Token inference cost and hardware utilization have reached the level of mainstream NVIDIA GPUs, verifying the feasibility of domestic computing power supporting the commercialization of hundred-billion-parameter models.
Regarding commercialization and ecosystem integration, the pricing strategy for GLM-5.3-Flash is highly competitive, with the official price set at only 十分之一 of GLM-5.3, further dropping to 四十分之一 of Opus 4.8 during the launch period. The model is currently open to the global community, seamlessly integrates with development platforms such as ZCode, distributes limited daily trial cards through the GLM Coding program, and simultaneously opens enterprise-level API interfaces for developer access.
KEY HIGHLIGHTS
Key highlights
Zhipu launches the world's first open-source native multimodal large model GLM-5.3-Flash, with a total parameter count of 320B and 18B activated parameters.
The model achieves a comprehensive intelligence index score of 57分, with code generation and logic debugging capabilities matching those of Claude Opus 4.8.
Adopts a hybrid sparse and linear attention architecture along with IndexPool technology, significantly reducing VRAM usage and latency for long-context processing.
Inference services are fully adapted to domestic AI chip clusters. Through EPD decoupled architecture and memory optimization strategies, end-to-end performance is improved by 3倍.
Priced at only 十分之一 of GLM-5.3, costs during the limited-time launch period drop to 四十分之一 of Opus 4.8. It is fully open-source and offers enterprise API access.