Qwen3.8 Release: 2.4T Parameter Open-Weight Model

Qwen3.8 Release: 2.4T Parameter Open-Weight Model

Qwen3.8 Introduces 2.4 Trillion Parameter Open-Weight Model

Alibaba has announced the launch of Qwen3.8, a massive large language model (LLM) featuring 2.4 trillion parameters. The model is designed to compete with leading frontier AI models and is scheduled to be released as an open-weight model, significantly expanding the availability of high-capacity models for the open-source community.

Alibaba claims that Qwen3.8 is one of the most powerful models available today, stating it is "second only to Fable 5" in terms of capability. A preview version, Qwen3.8-Max-Preview, is currently available for testing via Alibaba’s Token Plan, Qoder, QoderWork, and the Qwen chat interface.

Strategic Shift Toward Massive Open-Weight Models

The release of Qwen3.8 signals a shift in strategy among Chinese AI labs, moving from "value" models toward "intelligent, huge and slow" models. This trend is highlighted by the simultaneous competition with other high-parameter releases, such as Moonshot AI's Kimi K3, which features 2.8 trillion parameters and is slated for a Hugging Face release by July 27.

Industry observers note that this competition accelerates the availability of frontier-level capabilities in open-weight formats, potentially challenging the dominance of closed-source models from providers like OpenAI and Anthropic.

Community Insights and Technical Considerations

Technical discussions surrounding the Qwen3.8 announcement highlight several key points regarding deployment, performance, and accessibility:

Hardware and Inference Challenges

Due to the 2.4T parameter count, the full model is inaccessible to most small entities and consumer hardware. Community members have expressed a strong desire for:

  • Smaller Model Variants: Requests for 7B, 14B, and 20B parameter versions to enable local execution.
  • Optimized Quantization: A demand for A3B quants to make local inference more practical on consumer-grade GPUs.
  • On-Device Efficiency: A preference for models similar to the 27B class (e.g., Bonsai 27B) that balance a modest memory footprint with high speed and capability.

Performance and Censorship

While some users report high satisfaction with previous Qwen iterations (such as the 3.6 27B), others have raised concerns regarding the model's alignment and reliability:

"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."

Additionally, some users have characterized Qwen models as "benchmark princesses," suggesting that their performance on standardized tests may not always translate to real-world utility compared to competitors like Kimi.

The "Fable 5" Benchmark

The explicit mention of Fable 5 as the only model superior to Qwen3.8 underscores the current perceived gap between open-weight models and the absolute frontier of proprietary AI. The community views the eventual arrival of an open model that can surpass Fable 5 as a major milestone for the ecosystem.

Sources