inclusionAI/Ling-flash-2.0

Ling-flash-2.0 is a Mixture-of-Experts language model built under Ant Group's Ling 2.0 architecture, pairing 100 billion total parameters with a sparse activation strategy that fires only about 6.1 billion parameters per forward pass. This design lets the model punch far above the weight class of typical dense models, delivering reasoning and generation quality comparable to dense models upwards of 30–40 billion parameters while keeping inference more economical. The model is particularly tuned for complex, multi-step tasks: strong performance in mathematical problem-solving, intricate code generation, and frontend development reflects its training on diverse, high-quality data and its engineered expertise routing.

The model was trained on over 20 trillion tokens of curated data and refined through supervised fine-tuning followed by multi-stage reinforcement learning—a pipeline that shapes both factual accuracy and response style. Benchmark results cited from official documentation show competitive scores on multi-disciplinary knowledge reasoning (GPQA-Diamond, MMLU-Pro), advanced math benchmarks (AIME 2025, Omni-MATH), and coding evaluations (LiveCodeBench, CodeForces-Elo), positioning Ling-flash-2.0 as a capable generalist that also excels in technical domains. Released under an MIT license, the model is open-sourced and available in GGUF quantized formats for local deployment, with API access also provided through SiliconFlow for hosted inference.