China's LLM Spring: Alibaba, Zhipu, ByteDance Launch New Models—Is a Price War Coming?

China’s AI scene in April is buzzing like New Year’s Eve.

Alibaba launched three models in one week. Zhipu’s GLM-5.1 claims to be the world’s first open-source model for 8-hour continuous autonomous work. ByteDance quietly released their full-duplex voice model. Add in Tencent’s Hunyuan 3.0 teased earlier this month, and pretty much every major Chinese AI lab is in motion.

Here’s what’s interesting. I counted 12 new models released in the first 20 days of April—one every 1.6 days. That’s denser than the 2023 LLM boom.

Let’s start with Alibaba. Qwen3.6-Plus positions itself as the next-gen flagship for Agentic Engineering. MIT license open source—clearly gunning for Meta’s Llama developer mindshare. I tested its Agent capabilities, and tool-calling accuracy is noticeably better than 3.5. But honestly, what caught my eye was the pricing: 0.5 yuan per million input tokens, 2 yuan per million output. That’s nearly 90% cheaper than GPT-4 Turbo.

Zhipu’s GLM-5.1 took a different path. Focused on long-duration autonomous work, claiming it can work 8 hours without human intervention. I tested it on a complex data analysis pipeline—it broke down tasks, executed, corrected errors, and kept going. Still had some hallucinations, but overall reliability is way up from the previous generation.

ByteDance’s full-duplex voice model surprised me most. Not because of technical sophistication, but because of smart use-case targeting—customer service, education, companionship. Full-duplex means AI can listen and speak simultaneously, unlike current voice assistants that take turns. This interruption-friendly interaction is a qualitative leap for user experience.

But is the price war really here? My take: not fully, but the gunpowder is in the air.

Alibaba’s pricing is clearly loss-leading—cheap to gain market share and developer ecosystem. Zhipu plays premium, but pushes free tiers too. ByteDance hasn’t announced pricing yet, but given their track record, it won’t be expensive.

My feeling is that Chinese LLMs are transitioning from parameter competition to engineering deployment. Previously everyone compared how many billion parameters, now it’s does it work well in real scenarios and is it affordable? A healthy shift.

But there are concerns. Once price wars begin, smaller players may exit quickly. What remains will likely be the cloud giants with infrastructure backing. Good for developers—cheaper models, more choices. But tough for startups whose room gets squeezed further.

This is building bridges openly while crossing rivers secretly. On the surface, it’s new model releases. Underneath, it’s a fight for the AI infrastructure entry point of the future.

Final question: Are you using more Chinese or overseas models these days? As the gap narrows, this choice gets more interesting.