Alibaba's Qwen3.8-Max is out of preview, with an Arena ranking. It still has no independent benchmark score.

When [Alibaba](#/company/alibaba) first showed Qwen3.8-Max in July, it [claimed a rank of second globally with no benchmark table to back the claim up](#/article/alibaba-qwen38-max-preview-no-benchmarks). That gap has partly closed. Alibaba took the model generally available on August 3, publishing specifications, API pricing, and — for the first time — Arena.AI leaderboard placements it says put the model fifth on Text Arena and second on Vision Arena.

The specs: 2.4 trillion total parameters in a sparse mixture-of-experts architecture, with a hybrid attention mechanism that activates roughly 95 billion of them per pass. The [context window](#/dictionary) runs up to 1 million tokens. API access opened the same day on Alibaba Cloud's Model Studio at $2 per million input tokens and $6 per million output tokens, with cached input reads at $0.25 per million. Open weights — a first for a Qwen-Max-class model — are scheduled roughly a week after the API launch.

Alongside the model, Alibaba launched QwenWork, an agentic workplace product it positions against [Tencent](#/company/tencent)'s WorkBuddy, [Moonshot AI](#/company/moonshot)'s Kimi Work, Anthropic's Claude Cowork, and OpenAI's ChatGPT Work — four labs now shipping a near-identical product category within weeks of each other. Alibaba's own capability claim for the underlying model centers on a 16-day autonomous software-engineering project it says Qwen3.8-Max completed without human intervention; the company hasn't published the project's code, task specification, or a way for an outside party to verify what "autonomous" meant in practice, so that figure belongs with Alibaba's other self-reported claims rather than the independently checkable ones.

Deployment reality is narrower than the 2.4-trillion-parameter headline suggests. Running the flagship checkpoint locally requires infrastructure well beyond what most teams have; the practical self-hosting option once open weights ship is the smaller Qwen3.8-27B, also announced alongside the flagship. A developer choosing between the API and self-hosting is really choosing between the full model behind Alibaba's pricing and a materially smaller one whose benchmark scores haven't been separately published.

What "fifth in Text Arena" actually measures

Arena rankings are a real signal — thousands of people voting on blind, paired outputs isn't nothing — but it measures which answer people prefer to look at, not correctness under a fixed, adversarial test suite the way an independent [benchmark](#/dictionary) aggregate does. Alibaba does publish its own benchmark table, which shows Qwen3.8-Max at 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8's 84.6 on the same test but behind GPT-5.6 Sol's 88.8 in its highest-effort mode. Those three numbers, on the same test, are the closest thing to an apples-to-apples comparison available at launch. Max output is capped at 131,000 tokens per response — a ceiling worth knowing before assuming the 1-million-token figure describes what a single reply can contain, rather than what the model can read.

That table is Alibaba's own, not an independent lab's, which is exactly the gap the Scoreboard's rules exist to flag: a vendor citing a competitor's public score on a named test is at least checkable, but it is still the vendor choosing which test to feature. Readers comparing frontier models across vendors should weight vendor-published tables accordingly, whichever lab published them.

Where hands-on testing pushed back

Independent hands-on testing found real strengths: Qwen3.8-Max produced clean, functional front-end work — a self-referential landing page, a working Pokémon encyclopedia, a tourist map with real navigation — without the generic gradient-and-glow look common to AI-generated interfaces, and testers rated its agentic coding roughly comparable to Anthropic's Opus 4.5. It also caught the model taking a shortcut: asked to animate walking figures forming the text "Hello world, I'm Qwen," the model didn't choreograph the figures into letter shapes. It overlaid static text on the scene and arranged the figures separately, producing something that looked right at a glance without solving the actual spatial-reasoning problem the prompt implied.

The launch lands in a crowded field. [Moonshot](#/company/moonshot)'s Kimi K3 already carries an independent Artificial Analysis score of 57, DeepSeek's V4 Pro and V4 Flash are both measured, and Z.ai's GLM-5.2 has a published index score at a fraction of the price — all visible on this publication's own [Scoreboard](#/scoreboard). Qwen3.8-Max enters that field with a bigger parameter count and a longer context window than any of them, and no independent number to compare against theirs yet. Size and specification are not the same claim as measured capability, and this launch currently has the first without the second.

This publication's own Scoreboard reflects that gap directly: Qwen3.8-Max carries no score there, for the same reason Claude Opus 5 and DeepSeek V4 Pro sat unscored after their own launches — a vendor's own benchmark table and a crowd-voted leaderboard are not what that page measures. Until an independent aggregate publishes a number, the honest summary of Qwen3.8-Max's standing is Alibaba's own claim, a real but non-adversarial crowd vote, and one hands-on demonstration that at least one benchmark result doesn't mean what it looks like it means.