Open-Source Coding Models GLM-4.7 and DeepSeek V3.2 Compete
Two new open-source, MIT-licensed coding models, GLM-4.7 and DeepSeek V3.2, are emerging as top contenders for production use. A Novita analysis suggests GLM-4.7 is favored for its speed and stability in high-throughput applications, while DeepSeek V3.2 excels at more complex reasoning and code generation at a higher computational cost.
- Zhipu AI, the developer of GLM-4.7, originated from the Knowledge Engineering Group at Tsinghua University, one of China's top research institutions. The company is considered one of the "Six Tigers" of AI in China. - DeepSeek V3.2 utilizes a novel "DeepSeek Sparse Attention" (DSA) mechanism to improve computational efficiency, particularly in handling long contexts. It also employs a Mixture-of-Experts (MoE) architecture, which activates only a subset of the model's parameters for any given token, enhancing efficiency. - GLM-4.7 introduces a feature called "Interleaved Thinking," which allows the model to process information in a step-by-step manner before delivering an answer, aiming for greater accuracy and stability in complex, multi-turn conversations. - On the SWE-bench Verified metric, which assesses the ability to resolve real-world GitHub issues, GLM-4.7 scored 73.8%, slightly ahead of DeepSeek V3.2's 73.1%. However, DeepSeek V3.2 outperforms GLM-4.7 on other benchmarks like Terminal-Bench 2.0. - In terms of cost, DeepSeek V3.2 is significantly cheaper for both input and output processing. For example, one analysis found its output pricing to be about 5.2 times cheaper than GLM-4.7. - GLM-4.7 has a larger context window, accepting over 200,000 input tokens compared to DeepSeek V3.2's approximately 131,000. This can be an advantage for tasks requiring the model to process and reason over large amounts of information at once. - The MIT license under which both models are released is a highly permissive open-source license. This allows developers to freely use, modify, and distribute the software, including for commercial, closed-source applications, making it a popular choice for startups. - DeepSeek offers a specialized, high-compute variant named DeepSeek-V3.2-Speciale, which is designed for intensive reasoning tasks and has achieved gold-medal performance at the International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI). This version, however, does not support tool-calling and is intended more for research purposes.
