On January 8, on the day of the launch of Smart Spectrum, Tang Jie, professor of computer science at Tsinghua University, founder and chief scientist of Smart Spectrum, issued an internal letter announcing that the next-generation model GLM-5 will soon be launched. Tang Jie said that today is “an exciting day in Zhipu's life.” He did not directly respond to the big model company's business model controversy or give Zhi Spectrum's 2026 commercialization goals, but he emphasized that theories, technologies, or products that are truly “used by people” and can help more people are an important achievement of Zhi Spectrum in the pursuit of AGI. DeepSeek has had an impact on large model companies in China. Many people believe that DeepSeek's phenomenal success first impacted the intellectual spectrum ecosystem. They have almost the same academic research team attributes, and intellect also contributed a lot to the big model open source ecosystem. According to an internal letter, Zhi Spectrum completed the strategy set at the beginning of the year as scheduled in 2025, which is to release a “stable position” model in April, a “go to the table” model in mid-year, and release a Top 1 model at the end of the year. The strategy for this comprehensive return to basic model research is intelligent spectral response to the DeepSeek impact. On December 23, the intelligent spectrum base model GLM-4.7 was launched and open-sourced. Artificial Analysis showed that GLM-4.7 ranked first in China and tied with Claude 4.5 Sonnet for sixth in the world. In addition to the release of GLM-5, the internal letter also introduced the three technical directions that Smart Spectrum will focus on in 2026, including a new model architecture design, a more general RL paradigm, and exploration of continuous model learning and autonomous evolution. They all revolve around improving the capabilities of the basic model. As the basic model's capabilities improve, agents and domain models will eventually be combined with the basic model; even, AI doesn't necessarily mean that new applications need to be created. “The application of the big model also needs to go back to the principle of first nature.”

On January 8, on the day of the launch of Smart Spectrum, Tang Jie, professor of computer science at Tsinghua University, founder and chief scientist of Smart Spectrum, issued an internal letter announcing that the next-generation model GLM-5 will soon be launched. Tang Jie said that today is “an exciting day in Zhipu's life.” He did not directly respond to the big model company's business model controversy or give Zhi Spectrum's 2026 commercialization goals, but he emphasized that theories, technologies, or products that are truly “used by people” and can help more people are an important achievement of Zhi Spectrum in the pursuit of AGI. DeepSeek has had an impact on large model companies in China. Many people believe that DeepSeek's phenomenal success first impacted the intellectual spectrum ecosystem. They have almost the same academic research team attributes, and intellect also contributed a lot to the big model open source ecosystem. According to an internal letter, Zhi Spectrum completed the strategy set at the beginning of the year as scheduled in 2025, which is to release a “stable position” model in April, a “go to the table” model in mid-year, and release a Top 1 model at the end of the year. The strategy for this comprehensive return to basic model research is intelligent spectral response to the DeepSeek impact. On December 23, the intelligent spectrum base model GLM-4.7 was launched and open-sourced. Artificial Analysis showed that GLM-4.7 ranked first in China and tied with Claude 4.5 Sonnet for sixth in the world. In addition to the release of GLM-5, the internal letter also introduced the three technical directions that Smart Spectrum will focus on in 2026, including a new model architecture design, a more general RL paradigm, and exploration of continuous model learning and autonomous evolution. They all revolve around improving the capabilities of the basic model. As the basic model's capabilities improve, agents and domain models will eventually be combined with the basic model; even, AI doesn't necessarily mean that new applications need to be created. “The application of the big model also needs to go back to the principle of first nature.”