AI & ML 6 min read

Alibaba Previews Qwen3.8-Max at 2.4 Trillion Parameters, With No Benchmarks Published

Alibaba has opened preview access to Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it says trails only Anthropic's Claude Fable 5. What it has not published is a benchmark table, a model card, a licence, or the active-parameter count.

AI Tools Desk
AI & Developer Productivity Desk
Published 22 Jul 2026, 10:36 PM (SGT)
Share:
The Qwen3 logo — a purple geometric mark beside the wordmark on black The Qwen3 logo — a purple geometric mark beside the wordmark on black
Advertisement

Alibaba's Qwen team announced Qwen3.8 on 19 July 2026 and opened preview access to Qwen3.8-Max-Preview, a 2.4-trillion-parameter multimodal model, during the World AI Conference in Shanghai. The preview is live through Alibaba's Token Plan subscription and its Qoder and QoderWork coding platforms at a tenth of standard pricing, and the company says open weights will follow, without giving a date. Alibaba describes the model as competitive with leading frontier systems and, on its own account, second only to Anthropic's Claude Fable 5 among the models it tested. That ranking is the company's positioning rather than a measured result: Alibaba has not published a benchmark table to support it.

Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which is also named in this article. Readers should weigh that interest.

What Alibaba has confirmed, and what it has not

The confirmed facts are narrow. Qwen3.8-Max carries 2.4 trillion total parameters, accepts images, video and documents as well as text, and is the team's first multimodal model above one trillion parameters, according to Qwen developer Shuai Bai. Alibaba says it should outperform the previous Qwen3.7-Max on coding, full-stack development, data analysis and office workflows. The preview build is described as continuously evolving and may be changed or replaced when a production version ships.

Advertisement

The list of what is missing is longer, and every outlet covering the launch has noted it. There is no published benchmark table, no model card, no licence file, no open-weight release date, and no active-parameter count. The last of those matters most for anyone costing a deployment. Previous Qwen Max models have used sparse Mixture-of-Experts architectures, in which only a fraction of the network runs for any given token, although Alibaba had not published equivalent technical documentation for Qwen3.8-Max at the time of writing. On that lineage: Qwen3-235B-A22B activates 22 billion of its 235 billion parameters, and Qwen3-30B-A3B activates roughly three billion. Without the equivalent figure for Qwen3.8, the 2.4-trillion headline describes total size rather than the compute consumed per query, and says little about what the model will cost to run or to self-host. Dataconomy notes that earlier Qwen performance claims, including comparisons made for Qwen3.5, were never independently verified either.

The timing, and the open-weight question

The announcement landed two days after Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model whose weights are due on 27 July. The Decoder assesses Alibaba's release as aimed at Kimi K3's momentum. The timing invites that comparison, though Alibaba has not characterised the release in those terms and the inference is ours: two Chinese labs announcing multi-trillion-parameter flagships in the same week, at the same conference cycle, both promising open weights.

The open-weight pledge is the more consequential half. Alibaba has historically kept its Max-tier models closed and API-only, including Qwen3.7-Max in May 2026, while releasing smaller models such as the 397-billion-parameter Qwen3.5 openly. Following through on Qwen3.8 would break that pattern and give developers a model they can self-host and fine-tune rather than meter through Alibaba Cloud. Until a date, a licence and a repository exist, though, it remains a stated intention. For teams in Southeast Asia, where Alibaba Cloud is among the larger regional providers, the preview is straightforward to reach through Token Plan or Qoder; the harder decision is whether to build on a model that is explicitly labelled as subject to change, with no licence terms yet published to govern what a production deployment would rest on.

Key Takeaways

  • Alibaba announced Qwen3.8 on 19 July 2026 at the World AI Conference in Shanghai, with Qwen3.8-Max-Preview live via Token Plan, Qoder and QoderWork at 10 per cent of standard pricing.

  • The model carries 2.4 trillion total parameters and accepts text, images, video and documents; it is Qwen's first multimodal model above one trillion parameters.

  • Alibaba's claim that it trails only Claude Fable 5 is company positioning, not a published benchmark result.

  • No benchmark table, model card, licence, open-weight date or active-parameter count has been released; without the active-parameter figure, the 2.4-trillion count says little about inference cost.

  • Open weights are promised but undated, which would mark a departure from Alibaba's practice of keeping Max-tier models closed.

Advertisement
AI Tools Desk
AI & Developer Productivity Desk

AI Tools Desk tracks AI products, coding agents, model releases, and developer productivity tools for RECATOOLS.

View author profile → · Editorial policy

About this byline AI Tools Desk is a specialist RECATOOLS editorial desk focused on AI tools and developer productivity coverage. Articles are produced and reviewed under RECATOOLS editorial supervision. The articles listed here keep this byline. New coverage on these beats is published under the persona whose beat it falls in: Kenji Tanaka for AI and machine learning, developer tooling and cloud, and Amelia Wong for consumer-facing AI.

Corrections policy

Advertisement