New AI Models Optimized for Xiaomi TV Box Performance

• •Three new AI models optimized for Xiaomi TV Box S released.
• •Moon Beam 8B A1B runs at 7.2 tok/s, Bonsai 8B at 5.2 tok/s, and Qwen3.5 0.8B at 16.8 tok/s.
• •All models require a patched llama.cpp to operate correctly.

Three AI models have been released that run on the Xiaomi TV Box S (3rd Gen), which features 4× Cortex-A55 and ~1 GB of usable RAM. The models include Moon Beam 8B A1B, Bonsai 8B, and Qwen3.5 0.8B, each optimized for performance with different architectures. Moon Beam 8B A1B is an 8B mixture-of-experts model achieving 7.2 tok/s, while Bonsai 8B is a dense 1-bit model running at 5.2 tok/s. Qwen3.5 0.8B outperforms the stock model with a speed of 16.8 tok/s. All models require a patched version of llama.cpp from the smolmaxxing repository to function properly. The models have varying vocabulary sizes and performance metrics, with specific details on their architecture and memory usage provided in their respective documentation. Users are advised to follow the installation instructions carefully to ensure compatibility.