GLM 5.3 FlashX Deep Review: Zhipu’s High-Speed Multimodal Model, 200 Tokens/Second

Overseas brand of Zhipu AI Z.ai Launched on OpenRouter on September 18 GLM-5.3-FlashX— This is a high-speed variant of GLM-5.3-Flash, optimized for maximum inference speed and low cost. As the latest member of the GLM family, it inherits a hybrid sparse + linear attention architecture with 320B total parameters and 18B activated— packing multimodality, long context, and high throughput into a single model.

core competencies

GLM-5.3-FlashX’s most prominent label is “fast”: official benchmarks claim peak inference speeds of 200 token/sup to [value omitted], while OpenRouter’s real-world tests show P50 throughput of approximately 83 tokens/s and latency of around 2.32 seconds. For latency-sensitive scenarios—code completion, agent loops, and multi-turn dialogues—this speed delivers a tangible user experience improvement.

• Native multimodality: Both text and visual understanding are intrinsic model capabilities; image input is not an “add-on.”
• 1M-token ultra-long context: Capable of ingesting an entire technical document or a full execution trace of a long-horizon agent in one go.
• Hybrid sparse + linear attention: Only 18B of its 320B total parameters are activated, delivering large-model capability density at reduced computational cost.
• Budget-friendly pricing: Listed at $0.37 / $1.25 per million tokens; effective input cost is only ~$0.0987, with cache hit rate as high as 92%.

User experience/limitations

From a practical deployment perspective, GLM-5.3-FlashX exemplifies “fast to run—and affordable to run.” Its target use cases are clearly defined:Efficient programming, visual understanding, and long-horizon agent tasks. If you need a cost-effective, high-speed base model for heavy API usage, its value proposition is nearly unmatched.

But its limitations must also be clearly stated: as a Flash variant, it makes trade-offs in complex reasoning, deep mathematical capabilities, and code generation quality compared to the GLM-5.3 flagship base model—its output quality ceiling is lower than the flagship’s, an inevitable cost of “speed for depth.” Additionally, Z.ai is currently the only official provider for this model on OpenRouter, and the third-party hosting ecosystem remains limited.

Overall Score

Overall rating: 8.7/10

GLM-5.3-FlashX is a precisely positioned “high-speed multimodal” model: it brings large language models into high-concurrency, latency-sensitive production environments at extremely low cost and with outstanding throughput. If you’re looking for an affordable, fast, and Chinese-friendly general-purpose model to run agents or batch tasks, it’s worth trying. For authentic, in-depth reviews of AI tools, visit AI Dash—Discover the most useful AI tools.