GLM-4 is a new-generation pretrained model series released by Zhipu AI, and its open-source version is hosted in the THUDM/GLM-4 repository on GitHub. According to official materials, the project is an "open-source multilingual multimodal dialogue model," and a free trial is provided on Z.ai, while commercial model services are provided through bigmodel.cn.
Regarding project updates: on July 2, 2025, the GLM-4.1V-9B-Thinking series VLM was released; on April 14, 2025, the GLM-4-32B-0414 series models were released, with parameter scale expanded to 32 billion, including models with dialogue, reasoning, and Rumination capabilities; on June 18, 2024, a technical report was released; on June 5, 2024, the GLM-4-9B series open-source models were released.
Model introduction: The GLM-4-32B-0414 series has 32 billion parameters, and its performance can be compared with the OpenAI GPT series and the DeepSeek V3/R1 series, and it supports local deployment. Its base model, GLM-4-32B-Base-0414, was pretrained on 15T of high-quality data (including a large amount of reasoning-type synthetic data), laying the foundation for subsequent reinforcement learning scaling; in the post-training stage, human preference alignment is adopted, and rejection sampling and reinforcement learning are used to improve instruction following, engineering code, and function calling capabilities. It performs well in engineering code, artifact generation, function calling, search-based question answering, and report generation, and on some benchmarks (such as code generation or specific question-answering tasks) it can be comparable to larger models such as GPT-4o and DeepSeek-V3-0324 (671B).
GLM-Z1-32B-0414 is a reasoning model developed based on GLM-4-32B-0414. Through cold start, extended reinforcement learning, and further training on mathematics, code, and logic tasks, it significantly improves mathematical ability and complex task-solving ability, while introducing general reinforcement learning based on pairwise ranking feedback to enhance general capabilities.
GLM-Z1-Rumination-32B-0414 is a deep reasoning model with rumination capability (benchmarked against OpenAI's Deep Research). Unlike typical deep thinking models, it can think more deeply and for longer to solve more open and complex problems (for example, writing a comparative analysis of AI development in two cities and future planning). The model is trained through end-to-end reinforcement learning scaling, responses are scored by real answers or scoring criteria, and it can use search tools during the deep thinking process to handle complexMiscellaneous tasks, with notable improvements in research-oriented writing and complex tasks.
GLM-Z1-9B-0414 is a 9B small model trained with the above techniques, performing excellently on mathematical reasoning and general tasks, ranking among the top in overall performance among open-source models of the same size, achieving a balance between efficiency and effectiveness in resource-constrained scenarios, and providing a choice for lightweight deployment.
GLM-4-9B-0414, due to its smaller model capacity, did not undergo the same agent capability enhancements as GLM-4-32B-0414, but was mainly optimized for scenarios requiring large-scale batch processing (such as translation tasks).
Context and deployment: The model's native training context is 32K. When the total input + output length of a request may exceed 32K tokens, it is recommended to enable YaRN for better extrapolation performance; for the GLM-Z1 series, YaRN (Rope Scaling) can be considered when the input length exceeds 8,192 tokens. Most requests do not require modification if they do not exceed the native context length. The model implementation has been merged into the relevant repositories of vLLM, transformers, and llama.cpp.
Fine-tuning: The repository finetune/README.md provides information on the computational resources required for model fine-tuning and example scripts. The official team provides a simple fine-tuning example command; the script also supports visual tracking with SwanLab, and training logs can be viewed on the SwanLab visualization dashboard.
Prompt implementation: When using the transformers library's apply_chat_template to build prompts, the System Prompt restrictions differ across GLM-4-0414 models—GLM-4-32B-Base-0414 is a base model with no chat template; for GLM-4--0414 / GLM-Z1--0414, when tools are provided, apply_chat_template will fill the tools into the fixed template of chat_template, generating a separate system message with tool bindings placed at the first position of the message list (messages[0]), with the originally passed messages shifted back by one position overall. GLM-Z1-Rumination-32B-0414 does not support custom system prompts or custom tools; its tools and system fields will be ignored by apply_chat_template; using this model requires an external search engine or a custom retrieval API, and it supports four tools in totaltools: search (executes a search query and returns results; the parameter query is the search query string, using English except for Chinese proper nouns), click (clicks a link in the search results and jumps to the corresponding page; the parameter link_id is the link ID), open (opens a specified website and retrieves the content of any website via URL; the parameter url is the target web address or domain name), finish (completes the task, no parameters). In its fixed chat_template template, the thinking process uses English; if it needs to be changed to another language, the Language Used section in <Important Configuration> must be modified (currently supports Chinese and English).
Evaluation results: The GLM-4-0414 series sampled nearly 500 test cases each on SimpleQA and HotpotQA, providing basic search and click tools for all models, keeping other settings consistent, and taking the average of 3 runs. Among them, Moatless v0.0.3 uses the parameters response_format="react", thoughts_in_action=False, max_interations=30, with no retries for failed trajectories, and the others use default settings; Agentless v1.5.0 uses BGE as the embedding model and FAISS for similarity search, and to accelerate patch validation while maintaining performance, the single-instance run timeout was changed from the default 300s to 180s; OpenHands v0.29.1 did not use YaRN context extension, limiting runs to at most 60 iterations and summarizing history to prevent exceeding the 32K context, with the summarization configuration being llm_config="condenser", keep_first=1, max_size=32, and no retries for failed trajectories.
Citation: If you find the work helpful, you may cite the paper "GLM-4 series: Open Multilingual Multimodal Chat LMs | Open-Source Multilingual Multimodal Dialogue Models".
Other details (such as the complete model list, benchmark score tables, download links, and licenses, etc.) are not yet provided.
