Springboards says its new Flint model is designed to generate more varied answers than mainstream chatbots, targeting brainstorming work where repeated, high-probability responses can limit creativity.
Qwen 3 underpins Flint, which was trained to add randomness only at specific decision points rather than simply raising model temperature—a method Springboards says avoids the incoherence that broader randomness can cause.
25 LLMs tested 50 times each in the NeurIPS-winning “Artificial Hivemind” paper often converged on near-identical open-ended answers, reinforcing the startup’s claim that model homogeneity is widespread.
Advertisers and strategists testing Flint said it pushed ideas in different directions, though users also described the model as a prototype that can still fail when pushed too far.
OpenAI argues reliable models naturally converge on familiar responses and warns that stronger novelty can weaken coherence, underscoring the trade-off Flint is trying to navigate.