Zhipu unveils new model with 490 tokens/s inference speed

Chinese model developer Zhipu has released a new large model and is emphasizing an inference throughput of 490 tokens per second, according to a report by the news aggregator yeeyi. The number is the central technical claim in the item and appears to be the main selling point the company wants to highlight.

The report is titled around the new model and the 490 tokens/s figure, and it is brief. It does not spell out the model's specific version, parameter count, or release timeline. On the available evidence, this reads as a performance-focused move rather than a full account of a new capability stack.

Tokens per second is a core measure of how efficiently a model serves requests. It shapes how many users a given slice of compute can support, how quickly a single request responds, and therefore the cost and concurrency limits of an API service, which is why vendors frequently use it as a headline metric.

As China's model race shifts from raw capability toward deployment and cost, inference speed is gaining weight. With headline abilities converging, faster and cheaper serving is increasingly what wins enterprise customers and high-concurrency workloads.

Zhipu's decision to foreground 490 tokens/s fits that trend. High throughput usually comes from model architecture choices, inference-engine tuning, and hardware optimization working together, but the report does not describe how this figure was achieved, leaving the technical path to be confirmed officially.

For the wider industry, serving efficiency is becoming a second competitive axis alongside parameter counts and benchmark scores. When leading models are all good enough for most tasks, speed and cost become the variables that decide commercialization.

What to watch next: whether Zhipu details the model's version and how it will be made available, the hardware and concurrency conditions behind the 490 tokens/s figure, and whether that speed translates into real products and revenue. Until those details land, the number is best read as a directional signal.