Show HN: Flint – A 30B model fine-tuned for less repetition

In a landscape where artificial intelligence is increasingly fine-tuned for precision and accuracy, a new model named Flint has emerged, specifically designed to foster creativity rather than simply provide correct answers. Introduced by Springboards, Flint aims to address a growing concern in the creative industries: the outputs from current leading AI models are becoming monotonous and repetitive. While this convergence can be beneficial in fields such as law or accounting, it poses a significant challenge for creatives like writers, marketers, and strategists who thrive on diverse ideas and inspiration. Flint represents a shift in focus, prioritizing the generation of unique and varied outputs over mere correctness, marking a pivotal development in the intersection of AI and creativity.

The Divergence Model

Flint's design is rooted in the principles of divergent and convergent thinking, two essential components of the creative process. Divergent thinking involves exploring a wide range of possibilities, while convergent thinking focuses on narrowing those options down to a single solution. Traditional large language models (LLMs) have struggled with divergent thinking, often leading to a phenomenon known as "mode collapse," where the models converge on a limited set of responses to open-ended questions. This limitation can stifle creativity, as it results in outputs that lack variety and originality.

In contrast, Flint has been specifically engineered as a "divergence model." This means that it is trained to exhibit higher entropy—essentially, a greater degree of randomness—at crucial moments during the generation process. By doing so, Flint can produce a wider range of valid responses, allowing for less predictable and more imaginative outcomes. The model's ability to maintain a balance between creativity and coherence is a significant advancement, as it enables users to explore unconventional ideas without sacrificing the quality of the output.

The effectiveness of Flint is underscored by its performance on NoveltyBench, a benchmark that evaluates the distinctiveness of responses generated by AI models. Flint achieved a mean score of 7.47 distinct responses out of 10, significantly outperforming other state-of-the-art models such as Gemini 3.1 Pro, which scored 3.19, and GPT-5.4, which scored 2.54. This demonstrates that Flint is not merely a variation of its base model, Qwen3-30B-A3B, but rather a fundamentally different tool for creative tasks.

Performance Metrics

Further testing of Flint's capabilities reveals its unique position among AI models. When subjected to a similarity test using repeated prompts, Flint exhibited a mean intra-model similarity score of 0.721, indicating a broader range of responses compared to other models, which scored between 0.864 and 0.905. This lower similarity score suggests that Flint is less likely to produce repetitive outputs, allowing for a richer exploration of ideas.

In terms of inter-model similarity, Flint's average score of 0.672 further emphasizes its distinctiveness, making it the most varied model in comparison to others. Notably, Flint's performance on standard benchmarks such as MMLU-STEM and TruthfulQA MC1 indicates that its divergence tuning does not compromise its overall capability. Flint scored 78.9% on MMLU-STEM, matching its base model, while also achieving slightly higher accuracy on TruthfulQA and ToxiGen tests.

This balance between creativity and performance is crucial, as it suggests that AI can be designed to encourage innovation without sacrificing reliability. Flint’s approach challenges the prevailing trend of convergence in AI models, offering a new pathway for enhancing creative workflows.

Why it matters

The introduction of Flint signals a significant shift in how AI can be utilized in creative industries. By prioritizing output diversity and fostering a more expansive range of ideas, Flint serves as a valuable tool for creators who require inspiration rather than just answers. This model is not intended to replace existing larger models that excel in reasoning and knowledge work; instead, it acts as a complementary resource that enhances the creative process.

For technologists and creators alike, Flint represents an opportunity to rethink how AI can be integrated into workflows. Rather than relying on a single polished answer, Flint encourages the exploration of multiple starting points, enabling users to pursue various creative avenues. As AI continues to evolve, the development of models like Flint illustrates the potential for technology to support and amplify human creativity, rather than constrain it. The availability of Flint through the Springboards app provides an immediate avenue for users to experiment with this innovative approach to AI-driven creativity, inviting feedback and further exploration of its capabilities.