Alibaba Open-Sources Qwen3.8-2.4T-A95B, a 2.4-Trillion-Parameter Flagship Model

Alibaba’s Qwen team has open-sourced the weights for Qwen3.8-2.4T-A95B, a new flagship-scale model that the company says is its first fully open Max-class release for developers and researchers.

The headline numbers are notable even by frontier-model standards. Qwen3.8-2.4T-A95B is described as a 2.4-trillion-parameter model with 95 billion active parameters per token, suggesting a mixture-of-experts style architecture aimed at combining very large total capacity with more manageable inference-time activation. Alibaba also says the model supports a native context window of 260,000 tokens, expandable to 1.01 million tokens.

According to the company’s positioning, the model is designed for coding, office productivity, research assistance, and long-horizon agent workflows. That last category is increasingly important as model developers try to move beyond benchmark-driven chat performance toward systems that can complete multi-step tasks over extended periods with minimal supervision.

Benchmark and task claims

Alibaba says the model posts strong results on benchmarks including PaperBench, which focuses on paper reproduction tasks, and OSWorld-Verified, which measures computer-use ability. The release claims that in some metrics the model matches or exceeds several leading international models, though those comparisons should be read in the context of each benchmark’s setup and evaluation methodology.

The company also highlights long-duration autonomous work as a differentiator. In its description, Qwen3.8-2.4T-A95B can sustain complex coding tasks over multiple days and assist with reproducing and optimizing research papers. If independently validated, that would place the model in a strategically important category as AI labs race to improve reliability on extended agentic workflows rather than short interactive prompts alone.

Why the release matters

The significance of this launch is less about one benchmark number and more about access. Releasing weights for a model at this scale gives researchers and enterprise teams a rare look at a top-tier open model, especially one positioned as a flagship rather than a distilled or midrange variant. That could make Qwen more influential in the open model ecosystem, particularly among developers building coding agents, research tools, and long-context applications.

The deployment story also matters. Alibaba says the model is supported by mainstream inference engines and that ecosystem partners have used quantization techniques to reduce storage requirements, potentially widening the range of environments where it can be tested or deployed. In practice, hardware demands for a model of this class will still be substantial, but optimization support could help larger labs and infrastructure-rich enterprises experiment with it more readily.

The broader open-model competition

Qwen3.8-2.4T-A95B arrives as competition intensifies between closed frontier AI systems and increasingly capable open-weight alternatives. The model’s scale, long context window, and agent-oriented framing show how open releases are moving closer to the ambitions traditionally associated with proprietary flagship systems.

For the AI industry, the most important question is whether Qwen’s latest open release can translate its paper specifications and internal benchmark claims into consistent real-world performance. If it can, Alibaba will have strengthened its position not just as a major cloud and AI vendor, but as one of the most consequential contributors to the global open-source model stack.