A quantized variant of JoyAI's LLM Flash, this model runs in MXFP8 precision with the last 6 layers kept in BF16 — a common technique to preserve output quality at the tail end of the network while reducing memory footprint elsewhere. It handles long contexts up to 131K tokens, making it capable of processing lengthy documents in a single pass.