中文翻译

摘要: 126

Z.ai’s GLM-5.2 open weights model with 753 billion parameters released on June 16, which outperforms OpenAI’s GPT-5.5 in several coding problems and is about one-sixth of the cost (in dollars) is...

正文

Z.ai’s GLM-5.2 open weights model with 753 billion parameters released on June 16, which outperforms OpenAI’s GPT-5.5 in several coding problems and is about one-sixth of the cost (in dollars) is a major achievement in open-source AI.

Introducing GLM-5.2: Frontier Intelligence, Open Weights

– Significant improvements in coding and agentic tasks

– Strong long-horizon capabilities with a 1M context window

– Two levels of reasoning effort: GLM-5.2 (max) pushes the limits, while GLM-5.2 (high) strikes a strong…

pic.twitter.com/SjGPSVhePJ

You Might Be Interested In

Indian Government To Adopt Hybrid AI-Human Intelligence Model

MINI Introduces Locally Assembled Countryman C In India

Google Rolls Out Pixel June Drop

— Z.ai (@Zai_org)

June 16, 2026

In VentureBeat, GLM-5.2 scored 62.1 on the SWE-bench Pro benchmark to do real-world software engineering tasks, significantly higher than GPT-5.5’s 58.6. FrontierSWE measures long-horizon task performance, GLM-5.2 scored 74.4 percent, far surpassing GPT-5.5 at 72.6 percent and Anthropic’s Claude Opus 4.8 at 75.1 percent.

The cost efficiency is also remarkable; the GLM-5.2 API is priced at $1.40 per million input tokens and $4.40 per million output tokens, or $5.80 per million. GPT-5.5 is priced at $5 and $30 per million input and output, so it is an additional $35 per million, so GLM-5.2 is about one-sixth cheaper than GPT-5.5, and offers comparable or better coding capabilities.

Architecture and Technical Innovation

With 753 billion parameters in the model, it has a huge optimization called IndexShare, which uses the same indexer in every four sparse attention layers. This reduces the number of compute operations per token by 2.9 at a maximum context length of 1 million tokens, as discussed in the VentureBeat article.

GLM-5.2 also features a more powerful Multi-Token Prediction layer to support speculative decoding and enables users to extend the length of the tokens during inference by up to 20 percent. The

来源: Brave/newsx.com

采集时间: 2026-06-17 20:16:49