--- title: "China's open-weight model wave — flagship 大模型 you can download, not just rent" date: 2026-09-26 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-china-open-models language: en --- # China's open-weight model wave — flagship 大模型 you can download, not just rent > Across late August and September 2026, Alibaba, Tencent, Zhipu and ModelBest published open weights for trillion-parameter-class models — shifting the Chinese AI business model from metered APIs to downloadable infrastructure. For most of the past three years, China's best AI models lived behind an API paywall, rented by the token. In a burst of late-August and September 2026 releases, the country's largest labs inverted that model — publishing weights you can download, modify and self-host. The move is less about generosity than a new economic bet: give away the engine, charge for where it runs. ## What actually shipped - **Alibaba's Qwen3.8-Max** — released August 3, 2026, with open weights distributed the week of August 10. It is a 2.4-trillion-parameter sparse Mixture-of-Experts model that activates about 95 billion parameters per token, with a 1-million-token context window and native multimodal input. Alibaba calls it the first open-weight model at "Max" scale in the Qwen family, and it attaches a revenue-sharing clause for large commercial users. - **Tencent's Hy4 preview (混元 Hy4)** — open-sourced August 28 under the Apache 2.0 license. It carries 770 billion total parameters, 49 billion active per token, and a 1-million-token context. Tencent targets long-horizon coding, document-heavy office work and research, and ships both BF16 and FP8 weights on Hugging Face and ModelScope. - **Zhipu AI's GLM-5.3 and GLM-5.3-Flash** — released mid-August; the Flash variant ships under an MIT license with 320 billion parameters and is built to run end-to-end on domestic accelerator chips. - **ModelBest (面壁智能) and OpenBMB's MiniCPM5-2B** — open-sourced September 8, a 2-billion-parameter on-device model the partners say beats larger 4B-class open models across a battery of benchmarks. - **Alibaba's Qwen-Image-2.1** — a 7-billion-parameter open-weight image generation and editing model released September 21. The field is crowded because the incumbents kept shipping too: Moonshot AI's Kimi K3 (late July) and DeepSeek-V4-Pro (mid-August) were already part of the same open-weight cohort. ## The engineering idea: intelligence density, not size A common thread is **sparse MoE** — fat total parameter counts (trillions) paired with thin per-token activation (tens of billions). The payoff is flagship capability at a fraction of the serving cost of a dense model. Tencent lists Hy4 preview API pricing at **US$0.834 per million input tokens** on OpenRouter and **¥6 (about US$0.84) per million input tokens** on its own TokenHub, with output at ¥18 (about US$2.52). By comparison, Moonshot's Kimi K3 charges roughly US$3 per million input tokens — several times more. The second idea is "L4" openness: not just weights, but training data, recipes and toolchains. ModelBest did exactly that, releasing datasets and its reinforcement-learning framework alongside MiniCPM5-2B. ## Why it matters beyond China The wave lands on top of genuine demand. China's **National Data Administration** reported that daily AI token calls reached nearly **175 trillion in June 2026** — the highest of any country, or roughly 125,000 tokens per person per day. ByteDance's Doubao (豆包) was among the first models anywhere to cross 100 trillion daily token calls. At that scale, every fraction of a cent per token compounds into real money, which is precisely why cheap, self-hostable weights are attractive. For global developers, the practical upshot is optionality. A 770B or 2.4T model that runs on your own GPUs — or on domestic Chinese accelerators — means no API lock-in, no per-call bill, and no exposure to export-control surprises. Zhipu's claim that GLM-5.3-Flash runs fully on Chinese chips also links the open-model story directly to the domestic compute buildout covered separately. The business logic is a shift from "AI as a product" to "AI as infrastructure." Once the weights are free, labs compete on the ecosystem, the fine-tuning tooling and the cloud the model runs on — not on keeping the model behind a meter. ## Honest limitations - Release dates and specs above come from vendor announcements and English-language AI trade coverage (Alibaba Cloud blog, MLWires, RankLLMs, industry digests). We have not run the models ourselves, and vendor benchmark charts should be read as marketing until independently reproduced. - The "runs on domestic chips" claim for GLM-5.3-Flash is Zhipu's own statement; real-world throughput and operator coverage on non-Nvidia hardware vary widely. - The 175-trillion-token daily figure is a National Data Administration statistic (mid-2026) for aggregate Chinese usage, not a model-level breakdown, and the token-counting methodology is not public. - Revenue-sharing terms on Qwen3.8-Max are described by secondary outlets; the exact commercial thresholds were not published in the sources we reviewed. - Several releases are explicitly "preview" builds; APIs, licenses and prices can change after launch. ## What readers can do now - If you build with AI, test one of these open weights (Hy4 preview or Qwen3.8-Max) on your own stack this week — the download-and-self-host path is now viable for teams with a few high-end GPUs. - Re-benchmark your API bill: at Hy4's ~US$0.84 per million input tokens, workloads that were uneconomic behind a closed API may now clear the bar. - Watch the domestic-chip compatibility claims (Zhipu on Chinese accelerators) — if they hold up in independent testing, it changes where frontier models can legally and practically run. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-china-open-models Free to quote with attribution and a link to the original.