Alibaba's largest flagship model yet — a 2.4-trillion-parameter sparse mixture-of-experts model with a 1M-token context window, built for advanced coding, agentic long-horizon work, and in-depth research. Made widely accessible on August 3, 2026, with an open-weight release promised as the first Max-class Qwen to be open-sourced.
Last verified · 2026-08-03 · by Moe Ameen
Qwen3.8-Max is the largest and most capable model Alibaba's Qwen team has shipped, made widely accessible on August 3, 2026 after a preview earlier in July. Alibaba describes it as a 2.4-trillion-parameter model with a 1 million-token context window, positioned for advanced coding, real-world tasks, in-depth research, and long-horizon problem-solving. It is built on a sparse mixture-of-experts (MoE) architecture with a hybrid attention mechanism, so only a fraction of those 2.4 trillion parameters activate on any given token (reports put the active count on the order of 95 billion, though Alibaba has not fully detailed it) — that is what keeps a model this large economical to serve.
The headline positioning is coding and agentic autonomy rather than pure chat. Alibaba says the model can run coding tasks highly autonomously over long stretches, and pointed to an internal test in which it spent roughly 16 days building and improving an AI coding tool on its own — writing code, testing it, fixing errors, and refining the result with little human input. It is also multimodal, handling text, images, video, and documents, and Alibaba frames it as competitive with leading US frontier models: on its own benchmark reporting it edges ahead on several coding, multimodal, and engineering tests while trailing on some general-purpose reasoning ones. Treat that as vendor benchmarking until independent evaluations land — the direction (a coding-and-agentic-heavy flagship) is clear; the exact rankings are not yet settled.
Access is staged. You can use Qwen3.8-Max today through Qwen Chat and the Qwen API, and through Alibaba's discounted Token Plan, without waiting for the weights. Alibaba has said it will release the model's weights for public download and that this will be the first Max-class Qwen to go open-source — but at the time of writing the download and its exact license had not shipped, so confirm those on the official model card when they land. It arrived in a crowded window, launched close to Moonshot's 2.8-trillion-parameter Kimi K3, and is part of a fast Qwen cadence — so verify the current specs, pricing, and open-weight status on Qwen's official pages before you build on any single number.
Qwen3.8-Max is at its best on the thinking that happens before a single asset exists — the part a content workflow is usually weakest at. Point its 1-million-token context at everything you already have (a quarter of podcast transcripts, a research folder, a year of your own posts, a long product doc) and let it reason across the whole set at once: pull the recurring themes, rank the strongest angles, and draft an editorial plan with hooks and scripts. That long-horizon, agentic reasoning is exactly what a raw model does well and a scheduler does not. But the plan it hands back is still just words in a chat window — no video frame, no branded image, no post sized for a feed, and nothing published.
Kompozy is the execution engine that turns that plan into finished, on-brand content. Bring Qwen3.8-Max's angles and scripts into Kompozy, where your Persona Brief and banned-word filters rewrite them into your actual voice and fan each idea into formats the model can't produce — Persona and HeyGen avatar video, Clipped Shorts, brand-exact Carousels through HyperFrames, Quote Graphics, Photo Posts, plus native Text Posts, a Blog Article, and an Email Newsletter. Kompozy reframes every output to 9:16, 1:1, and 16:9, burns in branded captions, and schedules and publishes the whole set across nine platforms plus email and blog with Autopilot and a per-post review pipeline. Because Kompozy supports bring-your-own-key on the Founding tier, a team that has standardized on the Qwen API can keep Qwen3.8-Max as the planning and drafting brain while Kompozy owns the rendering, brand governance, and distribution the model was never built to touch. The model plans the campaign; Kompozy makes every piece and ships it.
Qwen3.8-Max is Alibaba's largest and most capable flagship model, made widely accessible on August 3, 2026. Alibaba describes it as a 2.4-trillion-parameter sparse mixture-of-experts model with a 1 million-token context window, built for advanced coding, agentic long-horizon tasks, and in-depth research. It is available via Qwen Chat, the Qwen API, and the Token Plan, with an open-weight release announced as coming.
They are closely related — Qwen3.8-Max is the flagship "Max" tier of the Qwen3.8 line, the specific 2.4-trillion-parameter model Alibaba made widely accessible on August 3, 2026 and positioned around advanced coding and agentic work. The broader Qwen3.8 announcement covered the same model family during its earlier preview. If you want the top flagship, that is Qwen3.8-Max.
Not yet at the time of writing. Alibaba said it will release the model's weights for public download and that Qwen3.8-Max will be the first Max-class Qwen to be open-sourced, but the download and its exact license had not shipped. For now it is accessible through Qwen Chat, the Qwen API, and the Token Plan. Confirm the license on the official model card when the weights are released.
Coding and agentic autonomy are its headline positioning. Alibaba says it can run coding tasks highly autonomously over long stretches — citing an internal test where it spent about 16 days building and refining an AI coding tool on its own — and claims it leads several coding and engineering benchmarks. Those are vendor claims pending independent evaluation, but the model is clearly aimed at long-horizon coding work.
No. Qwen3.8-Max generates and reasons over text (and reads images, video, and documents) but produces no finished video, images, or designs, enforces no brand voice, and publishes to no platform. To turn its output into on-brand posts across platforms, pair it with a content engine like Kompozy that renders the media and handles scheduling and publishing.