Show HN: Flint – A 30B model fine-tuned for less repetition (springboards.ai)

Springboards has introduced Flint, a groundbreaking 30-billion parameter AI model specifically designed to foster inspiration and creativity rather than provide standard answers. As industrial AI develops toward reasoning and precision, Flint targets creative industries where sameness in outputs—known as "mode collapse"—can stifle innovative thought. By enhancing output diversity without compromising accuracy, Flint aims to break the mold that conventional models have established, where they often generate similar responses to open-ended queries. Flint's training focuses on achieving higher entropy at critical points in its response generation, resulting in structured variations that yield more distinct and creative outputs. On the NoveltyBench benchmark, Flint achieved 7.47 mean distinct responses out of ten, far surpassing leading models like Gemini 3.1 Pro and GPT-5.4. Moreover, it maintains competitive accuracy on traditional benchmarks, proving that it can expand the range of creative responses while still delivering reliable results. This innovation signifies a shift in AI's creative potential, suggesting that Flint could serve as a useful complement to existing frontier models by enhancing creative workflows with more diverse ideas.