Z.ai update releases GLM-4.7 and positions it as the new flagship base model. According to official information, GLM-4.7 has significantly improved programming capabilities, complex reasoning and tool use compared to GLM-4.6, and has also been enhanced in scenarios such as chat, creative writing and role-playing; Model weights are publicly available in Hugging Face and ModelScope, and can be experienced directly in chat.z.ai.
On the capability side, GLM-4.7 emphasizes stability for "agent coding" and "long-link tasks": introducing a more granular thinking control mechanism, including inference before calling tools, and retaining inference content in multiple rounds of conversation to reduce information loss, thereby improving the consistency of multi-step execution. In terms of benchmark indicators, the official lists a number of comparative data such as SWE-bench Verified, SWE-bench Multilingual, and Terminal Bench 2.0 in the model card, and describes it as reaching the leading level of the open source camp.
GLM-4.7 has been gradually replaced as the default model of GLM Coding Plan, providing access guidance for common coding toolchains. The subscription page shows that the plan starts at $3/month. Since different regions, accounts, and tool configurations may affect the available models and experience effects, users still need to pay attention to quotas, call methods, and local computing power costs.
FAQs
Q: What model product is GLM-4.7?
A: The flagship large model version released by Z.ai focuses on programming, complex reasoning and tool calling, and provides open weights and online experience.
Q: What are the main improvements between GLM-4.7 and GLM-4.6?
A: The official emphasizes coding capabilities, complex reasoning, and toolchain execution stability, while enhancing the performance of chat and creative scenarios.
Q: Can GLM-4.7 be deployed locally?
A: Yes, the official provides an open weight download channel and gives deployment instructions under the mainstream reasoning framework, but the computing power and cost pressure are greater.
Q: What is the relationship between the GLM Coding Plan and GLM-4.7?
A: GLM-4.7 is already one of the default models for this subscription plan, providing access and configuration guidance for a variety of coding tools.
Q: What problem does GLM-4.7's "reserved thinking" mechanism solve?
A: It is used to reduce inconsistency between inference loss and inconsistency in multi-round tasks, and improve the controllability and completion rate of long-link proxy tasks.
