Cloudflare's decision model "Clef" playing Balatro (twitter.com)

Cloudflare has unveiled its new AI model "Clef," which successfully played and won the challenging game Balatro using NVFP4 quantization on a single RTX 5090 GPU, all while functioning fully offline. The accomplishment is notable as Balatro requires strategic thinking, and Clef lacks the chain-of-thought (CoT) planning capabilities seen in large language models (LLMs). Despite some suboptimal decisions during gameplay, Clef managed to complete the game in 19 minutes and 40 seconds, which is roughly three minutes faster than the previous GLM 5.3 Flash model, although it demonstrated slower response times averaging around four seconds as context size increased. The test highlighted several interesting aspects of Clef's performance, including a different economic strategy compared to GLM 5.3, with Clef utilizing a significant number of discards. This and other observations indicate potential areas for improvement in the model's decision-making capabilities and its harness for future iterations. The development of Clef is significant for the AI and machine learning community as it demonstrates advancements in offline model execution and the potential to tackle complex gaming scenarios without extensive resources.