China’s Zhipu AI and MiniMax simultaneously launched new large language models for coding on February 12, with both companies seeing sharp surges in their Hong Kong stock prices and valuations hitting new highs amid a bustling AI model release spree ahead of the Chinese New Year.
Dual Model Launches & Market Rally
Zhipu AI first announced the open-source of its GLM-5 model, and within half a day, MiniMax rolled out its M2.5 programming model. On the launch day, Zhipu AI’s stock price jumped as much as 25% at the opening, with a cumulative weekly gain exceeding 77% and its market value breaking HK$170 billion. MiniMax’s stock also rose over 20% on the same day, pushing its market capitalization above HK$180 billion, outperforming the broader market by a wide margin.
Zhipu AI’s market performance has been particularly striking since its IPO on January 8 at an issue price of HK$116.2 and an initial market value of HK$51.8 billion. Its valuation has more than tripled in just one month, a rare rally in Hong Kong’s AI sector.
The two firms timed their releases to February 12, the last working window for enterprises and developers before the Chinese New Year holiday. The launch allows for initial dissemination in the tech community ahead of the holiday and gives developers ample time to test and discuss the new models during the long vacation.
Frenzied AI Model Releases Before Holiday
The dual launches are part of a wave of AI model rollouts in China ahead of the Chinese New Year. Since late January, tech giants and AI startups have unveiled a string of new models, including Alibaba’s Qwen3-Max-Thinking, Moon-shot AI’s Kimi 2.5 and StepFun’s Step 3.5 Flash. On February 8, Alibaba’s Qianwen team submitted support code for Qwen 3.5 to Hugging Face, confirming the upcoming release of the new model.
ByteDance has also planned to launch Doubao 2.0 on February 14, which includes the full matrix of Seedream 5.0 and SeedDance 2.0. While DeepSeek’s V4 model’s launch during the holiday remains unconfirmed, the AI market has already seen unprecedented activity. Citibank noted on January 26 that 2026 marks a turning point for AI assistants from technical verification to mass adoption, with the Chinese New Year holiday emerging as a key window to test user habit formation.
Technical Breakthroughs of New Models
Zhipu AI’s GLM-5 is positioned for complex system engineering and long-range agent tasks, aiming to complete entire system engineering rather than just writing code—a model the firm calls agent engineering, the next stage of vibe coding, a concept echoed by Andrej Karpathy, former co-founder and AI chief of OpenAI. The previously mysterious Pony Alpha model in the open-source community was revealed as GLM-5’s anonymous test version.
GLM-5 boasts significant technical upgrades, with its parameter scale expanding from 355B (32B activated) in GLM-4.7 to 744B (40B activated) and pre-training data increasing from 23T tokens to 28.5T tokens. It features two core innovations: the Slime asynchronous reinforcement learning framework, which maintains strategic consistency in long-step engineering tasks, and the integrated DeepSeek Sparse Attention mechanism that cuts deployment costs while preserving long-text processing capabilities. GLM-5 ranks first among open-source models with scores of 77.8 on SWE-bench-Verified and 56.2 on Terminal Bench 2.0, and achieved a $4,432 balance in the Vending Bench 2 test, demonstrating long-term planning capabilities highly valued in the AI industry.
MiniMax’s M2.5, built natively for agent scenarios and benchmarked against Claude Opus 4.6, adopts a lightweight design with 10B activated parameters, offering advantages in video memory usage and inference efficiency. It supports full-stack programming for PC, App and cross-platform applications, and is optimized for productivity scenarios such as Excel, deep research and PPT. Developer feedback shows M2.5 processes daily coding tasks twice as fast as Claude Sonnet 4.5 with only 8% of the API cost, though the firm has not released its paper or Hugging Face weights.
Drivers Behind Stock Price Surges
The stock price rallies are not solely driven by new model launches. Zhipu AI announced a price hike of over 30% for its GLM Coding Plan, citing surging market demand and rising user scale and call volume, while increasing investment in computing power and model optimization. The firm’s MaaS platform annual recurring revenue surged from 20 million yuan to over 500 million yuan in 10 months after GLM-4.7’s release, a 25-fold increase.
For MiniMax, JPMorgan Chase initiated coverage on February 10 with an Overweight rating and a HK$700 target price, sending its stock up 10.7% on the day. JPMorgan highlighted MiniMax’s 73.1% overseas revenue share in the first three quarters of 2025, far exceeding domestic AI peers, and labeled it a “born global” AI enterprise. Its competitive pricing edge over international players like OpenAI, Anthropic and Google, combined with higher acceptance and payment capacity for AI tools in European and American markets, has become its core strength.
The flurry of model launches ahead of the 2026 Chinese New Year signals a new phase for China’s AI industry, with domestic firms shifting focus from parameter competition to developing commercially viable AI tools for developers.
