A single Reddit post condenses a month of work and a set of releases that might look like just another Hugging Face model drop. In practice, LLMFan46's package touches a raw nerve in local inference: the most recent architectures often land in forks before upstream support, and GGUF compatibility is never neutral.

At the center is LongCat-Flash-Lite-Sparse, a 69B-A3B model that adds sparse attention and a 1 million token context compared with the 256k of the dense variant. The author explains that support had to be created from scratch, both for Heretic and for a llama.cpp fork, because the upstream project does not support it. Anyone who wants to use the GGUFs must therefore go through llama-server.exe and llama-ui using a personal repository.

There are two variants: Uncensored Heretic with 9 refusals out of 100 and 0.0157 KLD, and Ultra Uncensored Heretic with 4 refusals out of 100 and 0.0779 KLD. Those numbers describe a more permissive behavior, but for a company they mean a different risk perimeter, not just a style preference.

The runtime is the real bottleneck

Depending on a personal fork changes the TCO of a self-hosted deployment. In an enterprise context, every component without upstream support is a maintenance cost and a potential exposure point. Downloading the weights is not enough: teams have to verify patches, manage updates, and accept that the path may diverge from the official one. The rest of the package confirms the trend. Qwen3.8-27B Ultra Uncensored Heretic with MTP has 3 refusals out of 100 and 0.0244 KLD, available in Safetensors, GGUF, NVFP4, NVFP4 GGUF, and GPTQ-Int4. Qwen3.5-122B-A10B arrives in GGUF with 8 refusals out of 100 and 0.0856 KLD. Qwen3-Coder-Next came from a Hugging Face user request, and Laguna-S2.1 adds an optional vision component, with the caveat that the vision part is far from perfect and can be excluded by not downloading the mmproj files.

The proliferation of formats has an infrastructural meaning. GGUF, NVFP4, and GPTQ-Int4 are not equivalent: they respond to different runtimes and GPUs, and a wrong standardization choice can force a team to change stack later. At the same time, the announced support for MoE Qwen3.5/3.6/3.8 models in the J-Wash Enhanced fork shows that the local ecosystem is organizing itself to fill the gaps left by the main projects.

The direction is clear: innovation in self-hosted inference moves faster than official channels. But speed has a price, and in this case it is paid in governance, updates, and trust in individual forks. It is not a verdict, but a condition to understand before adopting these models in production.