You are viewing a single comment's thread from:

RE: LeoThread 2025-11-09 20-32

in LeoFinance14 days ago

Part 5/10:

Community researchers also identified specialized tokens within the model’s weights—search begin, search end, think, and end think—enabling the model to perform real-time web searches and conduct internal, private reasoning processes before generating responses. These features, long-awaited by developers, facilitate advanced reasoning and updatable knowledge retrieval, making V3.1 a versatile tool capable of complex tasks in and outside traditional contexts.

Benchmark Triumphs and Limitations

V3.1 demonstrated impressive results across multiple benchmarks:

  • SVGBench: Near GPT-4.1 Mini, indicating strong reasoning in visual and structural tasks.