Skip to content
NewsModel release

Alibaba opens Qwen3.8-Max to global users and pledges open weights within a week

· by Pondero Newsdesk

The short version

Alibaba made its 2.4-trillion-parameter Qwen3.8-Max accessible globally on August 3 via Alibaba Cloud and the new QwenWork platform, and confirmed the weights will ship publicly within a week, marking the first Max-class Qwen model to go open-source.

Alibaba opens Qwen3.8-Max to global users and pledges open weights within a week

Alibaba on August 3 opened Qwen3.8-Max, its 2.4-trillion-parameter multimodal model, to global users and confirmed it will release the weights publicly within one week. That second commitment matters more than the API launch: every previous Max-class Qwen model stayed proprietary.

What happened

Alibaba opened access through two channels on August 3: Alibaba Cloud's Model Studio API and QwenWork, a workplace AI agent platform that entered public beta the same day on web and desktop, per South China Morning Post.

The model runs a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion active parameters at inference time, per Investing.com. Alibaba said it delivers lower inference costs than prior Qwen versions. The context window reaches one million tokens, covering full-length documents, long video streams, and multi-episode television series as inputs, per South China Morning Post.

Cited multimodal capabilities include rebuilding a software application from a screenshot, converting a 2D floor plan into an interactive 3D visualization, and processing live video streams. Inputs span text, images, video, and live streams, per AI Weekly.

Alibaba confirmed the weights will follow within one week and described Qwen3.8-Max as "the first Max-class Qwen model whose weights will be released as open source," per Investing.com. License terms have not been disclosed.

Alibaba's US-listed shares gained 5.1% to close at $122.25 on August 3, per Yahoo Finance. The company's Hong Kong-listed shares rose 6% to HK$124.00, per Investing.com.

Why it matters

The open-weights promise is the concrete change for AI operators, not the API access. Qwen3.5-Max and earlier Max-class releases stayed behind the Alibaba Cloud API. A team that wants to self-host a 2.4-trillion-parameter model now has a path, once license terms confirm commercial use is permitted.

The hardware requirement matters here. Running 2.4T parameters at practical latency requires significant GPU capacity, so the decision an operator actually faces is whether the weights enable fine-tuning and evaluation at smaller scale, not a full-deployment swap from a frontier cloud API. That distinction will be clearer once the weights ship and the community measures real inference costs.

QwenWork broadens the competitive picture for enterprise buyers. Per AI Weekly, the platform positions against Tencent's WorkBuddy and Anthropic's Claude Cowork. For teams already evaluating Claude Cowork or Microsoft 365 Copilot, QwenWork is now a Chinese-hosted alternative with tight integration to the Qwen model stack. Whether it supports custom agents or just exposes preset workflows is not yet specified.

On benchmark performance, South China Morning Post cited company claims placing Qwen3.8-Max "second only to Anthropic's Claude Fable 5." Investing.com reported benchmark testing showed the model "broadly competitive with leading U.S. AI models from OpenAI and Anthropic," outperforming on coding and multimodal tasks while trailing on some general reasoning evaluations. No independent benchmarking organization had published a full evaluation of Qwen3.8-Max as of August 3. These figures are vendor-framed until external assessments arrive.

The one-million-token context matches the ceiling on current Gemini 1.5 Pro and Claude models. At that length, the practical questions shift to latency, cost per million tokens, and whether Alibaba Cloud throughput can match production-scale demand from non-Chinese customers.

Context

The Qwen model line started as a family of smaller, open-source models. Alibaba moved toward proprietary releases for its highest-performing versions over the past year, then reversed that decision for Qwen3.8-Max. Whether the policy holds for subsequent Max-class models is not stated.

The QwenWork launch fits a pattern visible across Chinese AI providers: building application-layer products on top of their models rather than offering only raw API access. ByteDance's Doubao and Moonshot's Kimi have followed a similar trajectory. AI Weekly noted that Moonshot also widened developer access to Kimi K3 in the same week, putting two Chinese frontier providers in motion simultaneously.

The Qwen model family has built substantial developer adoption through its open-source lower tiers. That install base gives the open-weights release of Qwen3.8-Max a ready audience of teams already familiar with Qwen tooling and fine-tuning pipelines, which is an advantage proprietary-only releases do not carry.

What to watch next

Three milestones in the near term: the weights release itself, expected around August 10; the license terms, which determine whether commercial deployment is permitted; and the first independent benchmark results on standard coding and reasoning suites. The license structure is the actual decision gate for enterprise adoption. A research-only license produces a different market outcome than an Apache 2.0 or MIT release.

The QwenWork enterprise beta is a separate track. Alibaba's next quarterly earnings report is the earliest place to look for external signals on whether the platform has paying customers or meaningful usage outside China.

Sources