Technical Specifications and Model Architecture

inclusionAI has released Ling-3.1-flash-1019, an upgraded iteration of its previous Ling-3.1-flash model. According to documentation provided by inclusionAI, the new model utilizes a Mixture-of-Experts (MoE) architecture featuring 560 billion total parameters, with 25 billion active parameters per token. The model maintains a context window of 1 million tokens.

In addition to the chat-optimized version, the developer has released the Ling-3.1-flash-base model in BF16 format on Hugging Face. This base model is intended to provide a foundation for developers to conduct their own post-training processes.

Performance Benchmarks and Improvements

The release focuses on enhancements in reasoning and agentic capabilities achieved through further post-training optimizations. Based on data reported by inclusionAI, the model shows measurable gains across several key benchmarks:

  • Agentic Capabilities: The AA-Briefcase score increased from approximately 1400 to 1524 Elo, while performance on AutomationBench-AA rose from 62% to 68%.
  • Scientific Reasoning: The model demonstrated improvements in scientific benchmarks, with CritPt scores rising from 18% to 26% and SciCode scores increasing from 54% to 57%.
  • Intelligence Index: The model’s score on the Artificial Analysis Intelligence Index v4.3.2 rose to 44, up from 41 in the predecessor version.

These metrics suggest improved performance in multi-step task execution and complex reasoning workflows, according to the developer.

Community Reception and Implementation

Following the release, discussions on platforms such as Reddit indicate a focus on the model’s utility for specific high-volume tasks. Some users have noted the model’s efficiency in summarizing large quantities of documentation or code.

Technical users have noted a minor documentation error in the model’s README regarding SGLang implementation; the example currently points to the previous model path (inclusionAI/Ling-3.1-flash) rather than the new 1019 checkpoint. Users are advised to manually update the model path to inclusionAI/Ling-3.1-flash-1019 when implementing the new weights.

While the performance gains are documented by the developer, independent third-party verification of these benchmark improvements has not been provided in the current release announcement.