Zhipu AI Unveils GLM-5.1-HighSpeed, Achieving 400 Tokens/s and 6× Faster Generation

On May 22, 2026, Zhipu AI announced the GLM‑5.1‑HighSpeed variant, a fast‑track version of its GLM‑5.1 large language model.

Official measurements show the model can output 400 tokens in one second, delivering more than six times the throughput of the standard GLM‑5.1 and twice the speed of Google’s Gemini‑3.5‑Flash. The streaming output remains stable and coherent, with lower latency and higher throughput than typical large models.

The speed gain is not achieved by merely shrinking model parameters; Zhipu AI redesigned the underlying inference scheduler, refined the attention mechanism, and introduced multi‑dimensional sequence‑parallel computation. These changes accelerate processing while fully preserving the original model’s logical reasoning, copywriting, knowledge‑question answering, and long‑text handling abilities.

In the current global race for ultra‑fast large models, the launch of GLM‑5.1‑HighSpeed signals that domestically developed models have entered the first‑tier of real‑time inference, overtaking many overseas fast‑generation models.

As high‑speed LLMs become more widespread, low‑latency, high‑efficiency generation is expected to become a standard feature of future model products, potentially driving down token costs for end users.

Reference: https://docs.bigmodel.cn/cn/guide/models/text/glm-5.1-highspeed?webview_progress_bar=1&show_loading=0&push_animated=1&theme=light