Ling 3.1 Flash: 560B open-weight MoE, MIT license, 1M context (huggingface.co)

Ling-3.1-flash has been unveiled as an advanced hybrid reasoning model featuring 560 billion total parameters and a context window capable of handling up to 1 million tokens. This latest iteration improves upon its predecessor, Ling-3.0-flash, by enhancing performance across various real-world applications such as software engineering, office productivity, and healthcare. With benchmarks demonstrating strong results in elemental tasks via coding and general-agent capabilities, this model is set to elevate how AI interacts with complex multi-domain tasks. The significance of Ling-3.1-flash lies in its cutting-edge architecture, which includes 63 KDA and 9 gated MLA transformer layers, along with a sophisticated routing mechanism involving 512 routed experts. For developers, the introduction of LingCode—a tool designed for intricate long-horizon coding tasks—is expected to deliver notable performance gains. Furthermore, a comprehensive suite of optimizations, including speculative decoding and advanced batch processing capabilities, bolsters its efficiency. With open-source aspirations for LingCode, the AI/ML community anticipates significant contributions to enhanced coding agents and other impactful AI applications.