All stories
Feyospace S1 OBLITERATED weighted GGUF Ecosystem Community reports

Feyospace's 'uncensored' cyber model gets GGUF builds starting at 11 GB

mradermacher's weighted conversions package a refusal-modified 28B cybersecurity model in files from 11.0 to 22.5 GB. The card recommends a 16.9 GB Q4_K_M build; parent benchmarks do not establish GGUF performance.

By AITrending2 min read
Sources (4)

Why now The public weighted-GGUF repository was created October 15 at 16:20:19 UTC and last updated at 19:26:35 UTC. The reviewed snapshot lists actual downloadable quantized files; these dates describe this conversion repository, not the original Feyospace model's launch.

A smaller download for a refusal-modified cyber model

Community quantizer mradermacher has published weighted GGUF builds of OBLITERATUS’s Feyospace S1 modification, putting its cybersecurity-focused model into downloadable files starting at 11.0 GB. The October 15 conversion repository offers a range of sizes rather than one full-precision checkpoint.

The English card lists Q2_K at 11.0 GB, IQ4_XS at 15.4 GB, Q4_K_M at 16.9 GB and Q6_K at 22.5 GB. Its author labels Q4_K_M “fast, recommended.” That is the packager’s recommendation, not an AITrending speed or quality comparison. The weights’ disk size also does not establish total RAM or GPU memory required for a chosen context length.

What the ‘uncensored’ label means here

The parent, OBLITERATUS/feyospace-s1-obliterated, modifies Feyospace-v1.1-small, a roughly 28B model using the Qwen3.5 architecture. OBLITERATUS describes removing cybersecurity-specific refusal directions and markets the result as “uncensored,” aimed at authorized security research, penetration testing and CTF work.

Those are the author’s claims about a weight modification. They do not demonstrate that the model always follows instructions, preserves every capability or produces reliable security findings. The original Feyospace team, OBLITERATUS and the GGUF packager are separate contributors; this is not a newly announced official Feyospace architecture.

Choose a file, then check the runtime

The new repository identifies these as weighted/imatrix quantizations and links a separate repository for static conversions. Downloading a GGUF is only one part of local setup: the card does not pin a tested runtime version, platform or complete memory requirement for this model. It does not bundle an agent harness, tool permissions or target environments.

The OBLITERATUS card reports evaluations in BF16 on an A100-80GB. Those results are not tests of these GGUF files. AITrending reviewed the English cards and public file inventory, but has not run the conversions or reproduced the refusal and benchmark claims.

Related reading