Qwen 3.8 Release: 2.4T Parameter Open-Weight Model

Qwen 3.8 introduces a 2.4 trillion parameter open-weight model

Alibaba has announced the launch of Qwen 3.8, a large language model featuring 2.4 trillion parameters. The model is designed to compete with leading frontier AI models and is positioned as being second only to Fable 5 in terms of power. In a significant move for the open-source community, Alibaba has stated that Qwen 3.8 will be released as an open-weight model.

Availability and Access

While the full open-weight release is pending, a preview version titled Qwen3.8-Max-Preview is currently available for testing through the following platforms:

Alibaba’s Token Plan
Qoder
QoderWork
Qwen chat (offered for free)

Market Context and Competitive Landscape

The announcement of Qwen 3.8 appears to be part of a broader trend of high-parameter open-weight releases from Chinese AI labs, intensifying the competition between open-source and proprietary frontier models.

Competition with Moonshot AI and DeepSeek

Industry observers note that the Qwen 3.8 announcement follows a similar move by Moonshot AI, which announced Kimi K3, a 2.8 trillion parameter open-weights LLM scheduled for publication on Hugging Face by July 27. Additionally, users have highlighted the imminent release of the "final" version of DeepSeek 4, which is expected to compete at the level of Opus 4.8.

Comparison to Proprietary Models

Alibaba claims Qwen 3.8 is compatible with leading frontier models, though it explicitly acknowledges Fable 5 as the current benchmark leader. Community discussion suggests that the gap between open-weight models and proprietary ones (like those from Anthropic and OpenAI) is narrowing, with some users comparing the current landscape to the historical "OS wars" between Linux and Windows.

Community Insights and Technical Perspectives

Technical users and developers have provided several perspectives on the utility and performance of the Qwen series and the broader trend of massive open-weight models.

Local Inference and Model Size

There is significant demand for smaller, optimized versions of the Qwen 3.8 architecture. Users who rely on local inference for sensitive data have expressed hope for:

Smaller parameter sizes (e.g., 7B, 14B, or 20B versions).
Optimized A3B quants to maintain usability on consumer-grade hardware.
Models with modest memory footprints that prioritize speed and iterative capability over raw scale.

Performance and Censorship Concerns

While some users report that Qwen is superior to other Chinese models like GLM or Kimi, others have raised concerns regarding censorship and actual utility:

"Qwen is the most censored of the Chinese models in my testing... in my tests, existing Qwen models are not at the pareto frontier of any metric; DeepSeek V4 Pro is better, faster, and much cheaper than Qwen 3.7 Max."

Other users have characterized Qwen as a "benchmark princess," suggesting that its performance in standardized tests may not always translate to real-world utility compared to models like Kimi.

The Shift Toward "Intelligent, Huge and Slow"

Some observers have noted a strategic shift in Chinese AI development from "value" models to massive, high-intelligence models. However, this comes with trade-offs, as some users report that these larger models (including GLM 5.2 and Kimi 3) can be "extremely token hungry" and feel slower during actual use.