GLM-5 Is the Open-Weight Model Nobody Outside China Is Talking About

There is a tier of models that gets almost no English-language coverage despite being genuinely competitive, and GLM-5 is the clearest current example. Zhipu released it in May. It scores 85.8 on MMLU, 86.9 on HumanEval and 49.7 on SWE-bench Verified, with open weights, at $0.70 per million input tokens.

Those are not headline numbers. They are, however, better than the numbers most teams are actually paying frontier prices for, because most teams are running frontier models on tasks that do not need them.

The Fix That Matters

GLM-4.5 had a specific and well-known weakness: multi-step tool chains fell apart. It would call the first tool correctly, misinterpret the result, and then either call the wrong second tool or hallucinate that it had already finished. For agentic work that is disqualifying, and it is why GLM-4.5 stayed in the "interesting, cannot ship it" category.

GLM-5 fixes it. Not perfectly, but well enough that multi-step chains complete reliably in the five to eight step range that covers most real agent workflows. That single change moves it from a curiosity to a legitimate option for the cheap tier of a routing setup.

Bilingual Is Not the Same as Multilingual

GLM-5 is very strong in Chinese and English and mediocre in everything else. If your users are in those two languages, it is excellent value. If you need broad coverage, Qwen 3.8 handles 119 languages and is the better pick even though it costs slightly more.

This distinction gets flattened in most coverage. "Chinese model, good at Chinese" is the summary, and it misses that the Chinese labs have made different bets. Zhipu optimised depth in two languages. Alibaba optimised breadth. DeepSeek optimised cost. They are not interchangeable.

Practical Notes

• Context: 256K, with the usual middle-degradation caveat.
• Weights: open, and the model is small enough that self-hosting on a multi-GPU node is realistic.
• API latency: noticeably higher from outside Asia. If you are serving European or American users, self-host or use a Western inference partner.
• Vision: 70.2 on MMMU. Present, not strong.

Why It Stays Invisible

Distribution. GLM-5 is not in the default model list of the major agentic coding tools, it is not on most Western inference aggregators, and the documentation is China-first. None of that reflects model quality. It does mean that using it requires deliberate effort, and deliberate effort is exactly what most teams will not spend on a model that is not obviously better than their current default.

That is a reasonable position. It is also how genuinely good options stay underused for a year longer than they should.

Add it to a head-to-head on our model comparison tool and see how it lands against whatever you are running now.