Qwen4-Max Is Here: Alibaba's 2.4T-Parameter Flagship Cuts Agent Costs 62% — With a Catch

Photo: Hippopx (public domain). Illustrative image.
What happened. Alibaba launched Qwen4-Max on 16 November 2026, the flagship of its new Qwen4 family. The company pitches it not as a smarter chatbot but as a “24/7 AI teammate” for work: a model built to run long, multi-step jobs — coding, 3D design, manufacturing planning, finance, law and scientific research — and to hand back finished deliverables rather than drafts. It arrives a week after a Token Plan preview opened on 11 November; the full model is now generally available on Alibaba Cloud’s Model Studio.
The architecture. Qwen4-Max is a sparse mixture-of-experts model with a 2.4-trillion-parameter backbone, of which only 71 billion parameters activate per token, plus a further 328 billion N-gram Embedding parameters. Three changes do the heavy lifting, according to Alibaba’s announcement: Gated DeltaNet layers plus Qwen Sparse Attention compress history into fixed-size states for cheaper long-context attention; four parallel Gated Residual branches aim for more stable training; and N-gram Embedding adds capacity for local token patterns at little extra compute. Alibaba says the 2.4T-parameter training run progressed with no loss spikes throughout. On its in-house evaluations across knowledge, code, QA and STEM, the base model achieved the lowest bits-per-byte of five checkpoints tested — roughly 5–13% lower than Qwen3.8-Max-Base while activating 25% fewer parameters.
The cost claim is the headline. Against Qwen3.8-Max, Qwen4-Max uses fewer tokens on the same work: about 24% less on Alibaba’s Work360 evaluation at Xhigh effort and 35% less at High effort, and about 48–49% less on Terminal-Bench 4.0. Combined with lower API prices — particularly for cached input — Alibaba says the average cost per Work360 task drops from $4.43 to $1.67, a 62% reduction. Pay-as-you-go pricing on Model Studio is $1.50 per million input tokens and $4.50 per million output tokens. Two caveats matter. First, Work360 is Alibaba’s own benchmark, built on its own taxonomy of 26 top-level work domains, so the 62% figure describes Alibaba’s workload mix, not a universal price comparison. Second, Qwen4-Max scored 52.69 on Work360 against 47.74 for Qwen3.8-Max — a gain on its own test, which is expected. Independent benchmarks had not published scores at launch.
The demos lean into long-horizon work. In real production settings at SANF and Pacific Union, Alibaba says Qwen4-Max planned production lines, designed parts and verified machining, cutting the manual effort to program a single workpiece from about 475 minutes to 24 minutes. For pre-IPO legal due diligence it sorted 198 inconsistently formatted working papers into 14 folders and produced an 86-page report of over 60,000 Chinese characters, with every conclusion backed by citations. For investment research it invokes 90 MCP tools across market data. In healthcare it reviewed a six-day clinical course spanning dozens of lab reports and 1,433 medical images, and its four management recommendations matched the attending physician’s actual orders point by point. Two science showcases push furthest: a continuous 48-hour run that screened 2.7 million candidate molecules and directed an automated wet lab through four design–synthesis–testing loops to produce a physical sample of a novel coolant material, and a drug-discovery run targeting the NLRP3 inflammasome that made 454 tool calls and 3,239 computational evaluations across 24 candidate molecules, ending with a candidate whose estimated binding affinity is about 3.6 times the patent-derived reference — expert-endorsed, but still awaiting wet-lab validation. These are single showcases, not proof the loop completes reliably on any problem.
The catch. Qwen4-Max itself is not open-weight. The hosted flagship is available through QwenCloud (Token Plan or API), QwenWork, Qoder and Qwen Studio, and Alibaba says smaller Qwen4 models are “coming to open source” on the same lineage — those will be the ones to watch for self-hosters. Alibaba also open-sourced its Qwen-Live-Delivery-Plugin, the pattern behind the model’s staged delivery of long tasks: ship a usable first version early, then improve it through successive releases while the user can step in and redirect. On safety, Alibaba reports that on the Last Secure Code Bench, the function-level pass rate for both functionality and security rose from 22.0% for Qwen3.8-Max to 31.8%, and code secure at all three granularities — function, file and repository — rose from 8% to 14.7%. Nearly doubling is real progress, and a reminder that most generated code still is not secure by default.
Why it matters. Qwen4-Max is Alibaba’s bet that the next frontier is cheaper, steadier long-horizon work rather than a smarter chat window — a million-token memory, fewer tokens per task, and an agent loop that delivers early. If even half the efficiency holds on real workloads, it changes the maths for anyone running agents at scale. The model lands in a crowded week for flagship releases: Google’s Gemini Robotics 2.5 took robots to the factory floor, and DeepSeek opened application APIs to developers days earlier. For a plain-English explainer of the agent category Qwen4-Max is chasing, see What Is an AI Agent?.
Sources
- Qwen Team — “Qwen4-Max: Your New Teammate Has Arrived”, Qwen Blog (16 November 2026): https://qwen.ai/blog?id=qwen4-max
- AI First Look — “Qwen 4 Release: Qwen4-Max Is Out, API Price ($1.50 / $4.50), Specs and Early Tests” (16 November 2026, pricing and availability from Alibaba Cloud Model Studio documentation): https://aifirstlook.com/qwen-4/
- FreshSynapse — “Gemini Robotics 2.5 Hits 96% on a Mercedes-Benz Factory Line”: https://freshsynapse.com/articles/gemini-robotics-2-5-mercedes-benz-96-percent-2026-11-16/
- FreshSynapse — “DeepSeek Opens App APIs to Developers”: https://freshsynapse.com/articles/deepseek-api-application-tools-2026-11-15/
- FreshSynapse — “What Is an AI Agent? A Plain-English Guide to How AI Agents Work”: https://freshsynapse.com/articles/what-is-an-ai-agent/