Competition among China’s foundation model makers flared again this summer, with a fresh wave of model launches and technical one-upmanship taking center stage at the World Artificial Intelligence Conference (WAIC) in Shanghai.
This latest round saw a three-way showdown between China’s frontrunners: StepFun and Z.ai (formerly Zhipu AI), with tech heavyweight Alibaba returning to the fray.
On July 25, StepFun open-sourced its latest multimodal reasoning model, Step-3. That same day, Alibaba launched a new version of its inference model, Qwen 3. Just three days later, Z.ai released the latest generation of its flagship foundation model, GLM-4.5. But Alibaba wasn’t done. On the same day, it open-sourced Wan 2.2, a comprehensive multimodal toolkit supporting text-to-video, image-to-video, and unified video generation.
The rivalry was visible. On July 25, Alibaba declared Qwen 3 the “world’s strongest model.” But three days later, Z.ai’s GLM-4.5 overtook it on Z.ai’s own leaderboard, claiming third place globally among open-source models and knocking Qwen 3 down to ninth.
“We were all on edge,” a Z.ai employee told 36Kr. “Everyone on the algorithm team was glued to Qwen’s update. If the gap turned out to be too big, we’d be the punchline.” Tension eased only after GLM-4.5 pulled ahead across key benchmarks, including agentic evaluations.
The heat at WAIC mirrored a broader resurgence of competition across China’s AI sector in recent months.
Back in June, MiniMax staged a five-day product blitz that made headlines. Its open-source inference model M1 drew widespread attention with long-context handling and tool-use performance, while its video generation model, Hailuo 2, went viral for producing surreal content—such as diving cats.
In July, Moonshot AI’s open-sourced Kimi K2 swept 24 state-of-the-art (SOTA) titles in one shot.
Meanwhile, Baichuan AI and 01.AI were noticeably absent, both from WAIC and the current wave of model rollouts. Since the release of DeepSeek’s V3 and R1 models, most of China’s so-called “AI tigers” have gone quiet. Internal reshuffles and employee exits followed. According to job platform Maimai, by early July, around 41% of staff across six “tigers” marked themselves as “open to new opportunities.”
The outcomes of these model battles now carry existential weight, not just for regaining momentum, but for survival. Funding and commercialization prospects increasingly hinge on technical performance.
After six months of reputational drift and internal churn, many players are desperate for a defining moment to prove they still belong at the foundation model table.
But the path to redemption is steep. DeepSeek’s R1 model not only reset the performance bar, but showed that timing can matter as much as architecture.
Many still remember the sting. Kimi K2 had been in development since late 2024, slated as Moonshot’s showpiece for a mid-2025 release. But DeepSeek’s R1 beat them to launch. Moonshot scrambled, releasing a transitional model, Kimi K1.5, the same day as R1.
The rollout underwhelmed. It was only after Kimi K2’s stronger reception that Moonshot recovered some ground.
The K2 release also ruffled feathers, particularly at Z.ai, which was banking on dominance in coding and agentic performance.
To keep pace, Z.ai trained GLM-4.5 for nearly three months, abandoning its dense architecture for a mixture-of-experts (MoE) framework optimized for multi-agent tasks. “GLM-4.5 was initially conceived as China’s answer to Claude 4,” an industry insider told 36Kr. “But Moonshot kept everything under wraps so well, we only learned the technical details on K2’s release day.”
Caught off guard, Z.ai scrambled to accelerate training in the final month, ultimately elevating GLM-4.5’s coding and agentic capabilities to closely match those of Kimi K2. In doing so, it reclaimed a leading position among China’s “AI tigers.”
Once again, the clash of models has returned China’s AI contenders to center stage, evoking the high-stakes rivalries of 2023. But the dynamics have changed. The 2023 fixation on parameter counts and closed-source posturing has given way to a more transparent, technically grounded contest. Today, credibility is earned through both performance and openness.
Since DeepSeek’s breakout, open-sourcing models and publishing detailed technical documentation have become standard practice.
“The first users of a model are always developers,” said an AI app developer at WAIC. “If you don’t win over the developer crowd, it’s nearly impossible to build momentum.”
He likened open-sourcing to opening the front door to developers:
“It’s easy to reach them now as everyone’s on Hugging Face or GitHub. All you need to do is post your repo links.”
But open access is only the start. What matters next is what the model can do.
Despite a turbulent six months, most of the leading players still command far more capital and technical talent than the average startup. No one emerged as a definitive winner in this round of model warfare, but each had its moment:
- MiniMax has carved out a leadership role in AI video generation.
- Kimi K2 and GLM-4.5 each claimed SOTA honors across key benchmarks.
- Step-3 by StepFun leads the way in the still-niche but increasingly important multimodal segment.
There’s no clear champion, but all now hold visible positions on the global leaderboard. According to OpenRouter data, as of July 28, Kimi K2 ranked sixth worldwide by invocation volume. GLM-4.5 debuted in twentieth place the day it launched.
Meanwhile, Hailuo AI’s app surged in popularity after Hailuo 2’s release, with Diandian data reporting 110,000 downloads on July 22 alone.
Four of the original six “AI tigers” are now back in play, each having secured at least a temporary lifeline.
After five straight days of releases, including M1, MiniMax reportedly began IPO preparations. StepFun’s release of Step-3 was followed by a major funding round led by Shanghai State-owned Capital Investment and an ambitious RMB 1 billion (USD 140 million) annual revenue target, according to CEO Jiang Daxin.
Still, the competition is far from settled.
A growing consensus within the industry holds that China doesn’t need this many foundation models. Gaining ground now doesn’t guarantee a place in the long term. And the field isn’t limited to startups. Alibaba’s aggressive showing at WAIC made its point clear: in areas like multimodality, coding, and agentic reasoning—where smaller players can’t afford to tackle everything at once—tech giants can.
For those still in the race, the next round of eliminations may only be just beginning.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.