Home / AI / THUDM Unveils GLM 4: The 32B Powerhouse Rivaling AI Giants

THUDM Unveils GLM 4: The 32B Powerhouse Rivaling AI Giants

abril 14, 2025 10:47 pm

In the dynamic field of large language models (LLMs), researchers and organizations are tackling various significant challenges. These challenges include improving reasoning capabilities, offering robust multilingual support, and efficiently handling complex, open-ended tasks. Although smaller models are generally more accessible and cost-effective, they usually don’t match the performance of larger models. As a result, there’s an increasing focus on developing mid-sized models that balance computational efficiency with strong reasoning and instruction-following abilities.

The latest release from Tsinghua University, GLM 4, specifically the GLM-Z1-32B-0414 variant, effectively addresses these challenges. Trained on a significant dataset of 15 trillion tokens, GLM 4 is designed to provide dependable multilingual support and implements innovative reasoning strategies known as “thinking mode.” This places GLM 4 on par with other notable models like DeepSeek Distill, QwQ, and O1-mini and is distributed under the reputable MIT license. Remarkably, despite its moderate parameter size of 32 billion, GLM 4 delivers performance similar to much larger models like GPT-4o and DeepSeek-V3, which have up to 671 billion parameters, especially in reasoning-focused benchmarks.

From a technical standpoint, GLM-Z1-32B-0414 utilizes extensive high-quality training data, including synthetically created reasoning tasks, to enhance its analytical capabilities. The model incorporates advanced techniques like rejection sampling and reinforcement learning (RL) to improve its performance in agent-based tasks, coding, function calling, and search-driven question-answering tasks. Its “Deep Reasoning Model” variation further fine-tunes this by using cold-start methods combined with extended RL training, specifically aimed at complex mathematical, logical, and coding tasks. Pairwise ranking feedback mechanisms are used in the training process to improve the model’s general reasoning effectiveness.

An advanced version, GLM-Z1-Rumination-32B-0414, introduces an innovative approach called “rumination,” allowing for extended reflective reasoning to handle open-ended, complex queries such as comparative AI-driven urban analysis. This variant combines advanced search tools with multi-objective reinforcement learning, significantly boosting its utility in research-heavy tasks and complex retrieval-based scenarios. Supplementing these larger models, the GLM-Z1-9B-0414 version, with its 9 billion parameters, offers strong mathematical and general reasoning capabilities, showcasing the viability of smaller-scale models.