Release of Uncensored GGUF Models Including 1M Context LongCat-Flash-Lite-Sparse
Developer LLMFan46 has released a suite of uncensored GGUF models, highlighted by LongCat-Flash-Lite-Sparse, a 69B-A3B model featuring sparse attention and a 1M context length. This release also includes uncensored versions of Qwen3.8-27B, Qwen3.5-122B-A10B, Qwen3-Coder-Next, and Laguna-S2.1 with vision capabilities. The release provides the open-source AI community with highly capable, uncensored models that support advanced features like Multi-Token Prediction (MTP) and massive context windows. It enables local execution of large-scale models on consumer hardware through quantization formats like GGUF. Running the LongCat-Flash-Lite-Sparse GGUF requires using the developer's custom llama.cpp fork, as it is not supported in the upstream repository. The models preserve native Multi-Token Prediction (MTP) and Local Sparse Attention (LSA) to maintain generation speed and quality.
## BACKGROUND
Multi-Token Prediction (MTP) is a technique where a language model predicts multiple future tokens simultaneously to accelerate inference speed. Sparse attention, as opposed to dense attention, reduces computational complexity by focusing only on a subset of tokens, which is crucial for scaling context windows to lengths like 1 million tokens. GGUF is a popular file format designed for fast loading and running of LLMs on local hardware using tools like llama.cpp.