Five models in eight weeks — China's AI wave, and what the free-tier war is really about
Chinese labs shipped five frontier-class models in eight weeks, cementing the open-weight offensive reshaping AI pricing
The three lines
- Qwen3.8-Max, Kimi K3, DeepSeek V4-Flash, GLM-5.2, Seedance 2.5 — five models in eight weeks
- Most are open-weight and cheap to run — China keeps raising the bar for what costs nothing
- This wave is the common backdrop to ChatGPT's free tier and LG's open-source release
Key questions
- Is Chinese AI really moving that fast
- In release cadence, unambiguously. Alibaba's Qwen3.8-Max, Moonshot's Kimi K3, DeepSeek's V4-Flash, Z.ai's GLM-5.2 and ByteDance's Seedance 2.5 all shipped within eight weeks — the industry has started calling it the 'five models in eight weeks' phenomenon. On quality, assessments split between 'gap remains at the top end' and 'overtaken in specific domains' — most benchmarks are self-reported.
- Why do they release everything open-weight
- It converts constraint into distribution. Cut off from top-end chips, Chinese labs compensate by giving models away: free weights win developer ecosystems worldwide, offsetting a compute deficit with adoption. MoE architectures that slash inference cost serve the same strategy — cheap enough to run anywhere.
- What does it mean for everyone else
- The free baseline keeps rising. ChatGPT going unlimited for free users and LG open-sourcing a 750B model both read as responses to this wave. Users gain options; anyone selling AI faces a climbing threshold for what can still command a price.
One thread ties together this week's AI news. ChatGPT going unlimited for free users, LG open-sourcing a 750B-parameter model, the White House rushing a verification framework — behind each sits China's model barrage. Alibaba's Qwen3.8-Max. Moonshot's Kimi K3. DeepSeek's V4-Flash. Z.ai's GLM-5.2. ByteDance's Seedance 2.5. Five models in eight weeks — a cadence the industry has started calling by that name.
1. The strategy — converting compute deficit into adoption
The speed is, paradoxically, born of constraint. Cut off from top-end semiconductors, Chinese labs have converged on a two-part strategy. First, mixture-of-experts architectures cut inference cost — big-model knowledge at mid-model prices. Second, release the finished weights openly and win the world's developer ecosystem. What they cannot win in compute, they aim to win in adoption.
The effects show up first in pricing power. When a near-frontier model you can download and self-host refreshes every two months, paid API pricing erodes continuously. This page reads OpenAI's free-tier opening as a defensive response to exactly that erosion — a fight to hold the default position even where the margin has gone.
2. The eight-week roster
| Model | Lab | Note |
|---|---|---|
| Qwen3.8-Max | Alibaba | Flagship, agent-focused pitch |
| Kimi K3 | Moonshot | Long-context lineage |
| V4-Flash | DeepSeek | Low-cost fast inference, July 31 |
| GLM-5.2 | Z.ai (Zhipu) | Open-weight line |
| Seedance 2.5 | ByteDance | Video generation |
| (context) K-EXAONE 2.0 | LG (Korea) | 750B MoE, Apache 2.0 — same current |
The caveats are as clear as the roster. Nearly every benchmark cited for these models is self-reported, with no independent verification; and "open" spans very different license terms model to model. How to read a release headline against its license line is the standing reference "What 'open-source AI' actually means."
3. What remains
The next front is agents — most of the five lead their pitch with agentic performance, and Google's Gemini Robotics 2 extends the same race into embodiment. The regulatory blind spot sharpens in parallel: pre-deployment testing has nothing to grip once weights are public, the gap flagged in "Thirty days before launch." The consumer-side fallout of the free war is in "ChatGPT's free tier goes unlimited."