The Flint model (FlintForCausalLM) has been merged into llama.cpp via PR #30928, enabling local inference of this 39.8B parameter decoder-only mixture-of-experts transformer. It features 49 layers — 2 dense followed by 47 MoE layers — with approximately 2.5B active parameters per token and a hybrid attention mechanism comprising 35 sliding-window and 14 full-attention layers. The model is available in GGUF format on Hugging Face via Unsloth.