latentbrief
Back to news
General15h ago

Alibaba’s Qwen3.8-2.4T-A95B Model Goes Open Source on Amazon SageMaker

AWS ML Blog1 min brief

In brief

  • Alibaba’s Qwen team has released the Qwen3.8-2.4T-A95B model as open weights, marking a significant milestone in AI accessibility.
    • This powerful 2.4 trillion-parameter model is optimized for demanding tasks like multi-step coding and long-term planning.
  • By making its weights available, Alibaba empowers developers to customize inference behavior and avoid per-token fees, though it requires specialized GPU infrastructure to host.
  • The model features a hybrid architecture that combines linear attention with full attention, enabling efficient processing of up to 262K tokens-extendable to 1 million.
    • This design is tailored for agentic AI workloads, where models must maintain state across many interactions without relying on inefficient chain-of-thought methods.
  • The release includes support for Amazon SageMaker HyperPod, allowing seamless deployment using vLLM on high-performance instances like ml.p6-b300.
  • With this move, Alibaba joins other companies in democratizing access to cutting-edge AI models.
  • Developers can now experiment with a model capable of handling complex reasoning and tool usage, setting the stage for new innovations in AI capabilities and applications.

Terms in this brief

Hybrid architecture
A combination of different attention mechanisms in neural networks to improve efficiency and performance. In this case, it combines linear attention with full attention, allowing the model to handle longer sequences more effectively.
Chain-of-thought methods
A problem-solving approach where models simulate a sequence of reasoning steps, often used in complex tasks. This method can be inefficient, so the hybrid architecture avoids it by maintaining state across interactions instead.

Read full story at AWS ML Blog

More briefs