Deepseek, Zhipu, and MiniMax collectively announce new releases

Exciting news continues to emerge about China’s AI large models.

On the evening of February 11, Zhipu officially confirmed that the mysterious model “Pony Alpha,” which previously topped the popularity charts on the global model service platform OpenRouter, is Zhipu’s new model GLM-5. The new model is now available on the chat.z.ai platform.

On February 6, the global model service platform OpenRouter quietly launched an anonymous model codenamed “Pony Alpha.” Due to its strong coding capabilities, ultra-long context window, and deep optimization for intelligent agent workflows, it quickly attracted attention from the developer community and gained rapid popularity in overseas communities.

OpenRouter describes Pony Alpha as a “cutting-edge foundational model” with strong performance in programming, intelligent agent workflows, reasoning, and role-playing, emphasizing its “extremely high tool invocation accuracy.” This feature gives it a significant advantage in AI agent application scenarios, allowing developers to use tools like Claude Code to invoke the model for complex project development lasting several hours.

On January 8, Zhipu officially listed on the Hong Kong Stock Exchange. On the listing day, the company’s Chief Scientist and Tsinghua University Computer Science Professor Tang Jie sent an internal letter confirming that the next-generation base model GLM-5 “is about to be released,” and announced that starting in 2026, the company will “fully return to foundational model research.” Additionally, they established the Frontier Innovation Department X-Lab, focusing on architecture, learning paradigms, and continuous evolution.

Furthermore, DeepSeek has also updated its models. According to reports, multiple users have provided feedback that DeepSeek has undergone version updates on both web and app platforms, supporting a maximum context length of 1 million tokens. Last August, DeepSeekV3.1 extended its context length to 128K.

Currently, few models can push context to the million-token level, with Google’s Gemini series and Anthropic’s Claude Opus 4.6 among the first to achieve this.

DeepSeek’s V-series models are positioned as foundational models pursuing ultimate comprehensive performance. The V3 model, launched in December 2024, marks a significant milestone for DeepSeek, with its efficient MoE architecture establishing a strong foundation for overall performance. Since then, DeepSeek has rapidly iterated on V3, releasing enhanced reasoning and agent capabilities in V3.1, and officially launching the latest version V3.2 in December 2025. They also introduced a special version, V3.2-Speciale, focused on solving complex mathematical and academic problems.

Tech media The Information previously reported that DeepSeek plans to launch a new flagship AI model, DeepSeek V4, during the Lunar New Year in mid-February, which will have even stronger coding abilities.

Earlier this year, the DeepSeek team published two papers revealing two innovative architectures: mHC (Manifold-Constrained Hyperconnection), designed to optimize information flow in deep Transformers for more stable and scalable training without increasing computational load; and Engram (Conditional Memory Module), which decouples static knowledge from dynamic computation, using inexpensive DRAM to store factual knowledge and freeing expensive HBM for reasoning, significantly reducing long-context reasoning costs.

On the same day, news also came that MiniMax’s M2.5 model is about to be officially launched. Currently, the MiniMax M2.5 model is in closed beta testing within MiniMax Agent products overseas.

(Source: Daily Economic News)