Release of Uncensored GGUF Model Variants

A series of new, uncensored large language model (LLM) variants in GGUF format has been released by developer LLMFan46. The release includes several high-parameter models, notably featuring “Heretic” and “Ultra Uncensored Heretic” versions designed to minimize refusal rates during user interaction.

According to reports on Reddit, the release includes a significant update to the LongCat-Flash-Lite-Sparse model. This 69B-A3B model reportedly incorporates sparse attention mechanisms and supports a context length of up to 1 million tokens, an increase from the 256k context length found in the original LongCat-Flash-Lite. The developer noted that the sparse attention support required custom development for the llama.cpp framework. The uncensored variants of this model report refusal rates of 9/100 for the “Heretic” version and 4/100 for the “Ultra Uncensored Heretic” version.

Expansion of Qwen-Based Models

The release also features several models based on Alibaba’s Qwen architecture. These include:

  • Qwen3.8-27B: An “Ultra Uncensored Heretic” version with reported refusal rates of 3/100.
  • Qwen3.5-122B-A10B: An “Uncensored Heretic” version with reported refusal rates of 8/100.
  • Qwen3-Coder-Next: An uncensored version of the Qwen3-Coder model, reportedly developed following specific user requests on Hugging Face.

The developer also released a vision-capable version of Laguna-S2.1. While the vision capabilities are described by the developer as not yet perfected, the model can function as a standard text-only model if the vision-specific files are omitted.

Technical Implementation and Community Response

Due to the specialized architecture of the LongCat-Flash-Lite-Sparse model, users cannot utilize the mainline version of llama.cpp. Instead, the developer has provided a specific fork of the software on GitHub to support the model’s unique requirements. Additionally, an enhanced version of the “J-Wash” tool has been released, which includes support for Mixture-of-Experts (MoE) models within the Qwen3.5, 3.6, and 3.8 families.

The release has garnered attention on social media platforms, with some users expressing support for the availability of uncensored models, citing concerns over the content moderation policies implemented by major technology corporations.

While these releases represent a significant technical undertaking by an individual developer, the performance metrics and refusal rates provided are self-reported and have not been independently verified by third-party benchmarks.