Clef Flash 9B plays Snake in real time on an RTX 5080
A developer demonstrated that Clef Flash, a 9-billion parameter model quantized to Q4 and running locally on an Nvidia RTX 5080, can play Snake in real time. The model achieved a decision latency of 135ms using simple prompt instructions without any game-specific training or state hacking. This proof-of-concept highlights how small, quantized language models running on consumer hardware can achieve real-time, human-like decision latency for agentic tasks. It demonstrates that lightweight local models can be applied to dynamic environments and gaming without complex domain-specific fine-tuning. The 135ms decision latency matches the turn limit of Google Snake, enabling human-speed gameplay entirely through prompt-based decision-making. Clef Flash processed state and visual input zero-shot to determine the snake's next move on each game tick.
## BACKGROUND
Clef Flash is a 9-billion parameter open-weight decision model post-trained from Qwen3.5-9B to quickly turn state information into decisions. Quantization (such as Q4) compresses a model's weights into 4-bit representations, significantly reducing memory bandwidth and GPU VRAM requirements while enabling low-latency inference on consumer GPUs like Nvidia's RTX 5080.