Zhipu's GLM-5.3 HighSpeed spotted in ZCode at nearly 500 tokens/s
Zhidx reported on Oct 12 that Zhipu has quietly opened a gray-scale test of a high-speed flagship model, GLM-5.3 HighSpeed, inside its AI coding tool ZCode, available only to randomly selected ZCode subscribers with no public application channel. X user Leo said he was granted access and saw output above 490 tokens/s; for comparison, AI Ping lists DeepSeek V4.1 Flash's official API at 121.34 tokens/s and Zhipu's own GLM-5.3 API at 70.84 tokens/s, and the model is said to carry a 1M-token context window with up to 131,072 output tokens, with pricing undisclosed. Zhipu offered a 400 tokens/s GLM-5.1-highspeed API to selected enterprise users in May, attributing the gains to system-level work with TileRT across the inference engine, scheduling and infrastructure; it has also posted the natively multimodal GLM-5.3-Flash on Hugging Face (320B total / 18B active parameters), whose model card claims it beats GLM-5.2 on benchmarks and real-world workloads at one-tenth the price and approaches Claude Opus 4.8 on coding and agentic benchmarks.
If the speed holds once the model is generally available, it could cut the multi-round latency that dominates coding-agent workflows and make flagship inference speed a competitive axis.