--- title: "Alibaba's Qwen3-Max Crosses One Trillion Parameters — but Stays Behind a Paywall" date: 2026-10-03 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-qwen3-max language: en --- # Alibaba's Qwen3-Max Crosses One Trillion Parameters — but Stays Behind a Paywall > Unveiled at Alibaba's Apsara Conference in September 2025, Qwen3-Max is the company's largest large model (大模型) at over one trillion parameters and 36T training tokens — but it ships as a hosted API, not open weights. A model bigger than almost anything China has shipped was unveiled to a packed conference hall — then handed to developers as a service, not a download. The same company gives away hundreds of smaller models for free. This one, it keeps. ## What actually shipped At Alibaba's Apsara Conference in late September 2025, the Tongyi (通义) team unveiled **Qwen3-Max**, which it calls its largest and most capable large model (大模型) to date. The Base version carries **more than one trillion parameters** and was pretrained on **36 trillion tokens**, following the Qwen3 MoE design with a global-batch load-balancing loss. Alibaba describes a notably stable training run — no loss spikes, no rollbacks — and a 30% relative gain in training efficiency (MFU) over its prior Qwen2.5-Max base. Qwen3-Max ships in two flavours: **Qwen3-Max-Instruct** for general use, and **Qwen3-Max-Thinking** for explicit reasoning. The Instruct model is available through Qwen Chat and Alibaba Cloud's Model Studio API; the Thinking variant was still in training at the time of the technical write-up but already showing extreme results. ## The closed-flag contrast Here is the part easy to miss: unlike the hundreds of Qwen models Alibaba has open-sourced, **Qwen3-Max does not ship open weights.** The official materials point developers to the hosted API and the Qwen Chat web interface. That is a deliberate split. Alibaba has built the most-downloaded open model family on earth — over 300 models, 600 million downloads, 170,000 derivatives — yet it is keeping the crown jewel behind a paywall. At the same conference, Alibaba also showed Qwen3-VL and Qwen3-Omni, and reaffirmed a three-year, **RMB 380 billion (≈ US$52.5B / HK$410B)** commitment to AI and cloud infrastructure. The strategy is clear: open the ecosystem, charge for the apex. ## What the benchmarks claim Qwen3-Max-Instruct's preview ranked **third on the LMArena text leaderboard**, ahead of GPT-5-Chat. The formal release pushed coding and agentic scores higher: **69.6 on SWE-bench Verified** and **74.8 on Tau2-Bench**, the latter described by Alibaba as surpassing Claude Opus 4 and DeepSeek V3.1 on tool-calling. The Thinking variant is the headline-grabber. Augmented with a code interpreter and parallel test-time compute, Qwen3-Max-Thinking scored **100% on AIME 25 and HMMT** — two gruelling maths benchmarks — according to Alibaba. Those are vendor-reported numbers on tasks where tool use and extra compute swing the result, so read them as Alibaba's claim, not a verdict. ## Why "just scale it" is the thesis Alibaba's own blog title for the release is "Just Scale it." The argument is unfashionable in a year when everyone preaches efficiency: Qwen3-Max's gains came from simply making the model and its data bigger. The company leans on long-context training tricks — a ChunkFlow strategy said to deliver 3× throughput over context parallelism, enabling a **1-million-token training context** — and on fault-tolerance fixes that cut hardware-failure downtime to a fifth of the prior generation. Whether scaling alone still buys frontier performance is the live debate. Qwen3-Max's answer is a confident yes; the rest of the field is betting on reasoning and agents instead. ## Pricing and access Alibaba lists tiered API pricing for Qwen3-Max that rises with context length. For the shortest tier (≤32K tokens) reported rates run about **US$1.20 per million input tokens and US$6.00 per million output tokens**, scaling to US$3.00 / US$15.00 for 128K+ prompts. Pricing differs by region and deployment (international vs mainland China), so a production budget needs the console figure for your zone. ## Honest limitations Parameter count (">1 trillion"), 36T training tokens, MoE design, LMArena ranking, SWE-bench 69.6, Tau2-Bench 74.8, and the AIME 25 / HMMT 100% Thinking scores all come from Alibaba's official Qwen blog and the Apsara Conference coverage (Xinhua, Alibaba News). The "2.4 trillion parameters" figure that appears in some secondary outlets is **not** supported by Alibaba's primary materials, which state "over 1 trillion"; we used the primary figure. Qwen3-Max **weights are not open** per the official release — only API/Qwen Chat access is documented. The **US$1.20 / US$6.00** pricing is the ≤32K international tier as reported by model aggregators; longer-context and domestic tiers differ. The **RMB 380 billion** infrastructure commitment is from Alibaba's official newsroom. All benchmark numbers are vendor-reported and lack independent third-party audit at writing. ## What readers can do now - **If you want the strongest Qwen:** call qwen3-max through Alibaba Cloud Model Studio or Qwen Chat — there is no weight download, so plan for API dependency. - **If you need open weights:** stay on Qwen's open releases (Qwen3-235B and smaller) and treat Max as a capability ceiling to compare against, not a model you can self-host. - **If you benchmark:** wait for independent SWE-bench / Tau2 numbers before trusting the 69.6 / 74.8 claims in procurement decisions. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-qwen3-max Free to quote with attribution and a link to the original.