license: apache-2.0
base_model: OBLITERATUS/feyospace-s1-obliterated
base_model_relation: quantized
library_name: gguf
pipeline_tag: text-generation
tags:

• gguf
• llama.cpp
• quantized
• imatrix
• cybersecurity
• abliteration
• tool-use
• mtp
• speculative-decoding

Feyospace-v1.1-Abliterated-security-MTP-UD-Q4_K_S

A community GGUF quantization of OBLITERATUS/feyospace-s1-obliterated, intended for local cybersecurity research, code analysis and tool-assisted investigation.

The release recipe combines the upstream cybersecurity-focused model and abliteration with security-oriented importance-matrix calibration, a mixed-precision UD-Q4_K_S tensor layout, and retention of the model's multi-token prediction (MTP) head.

Model details

The upstream checkpoint is multimodal. This card covers the text GGUF workflow; image and video inference are not validated for this artifact.

Quantization

The documented workflow converts the source checkpoint to a full-precision GGUF, collects activation importance statistics from Feyospace, and quantizes it with an explicit per-tensor type recipe.

UD-Q4_K_S identifies a mixed-precision layout. It does not mean that every tensor uses four bits. mixed-precision layout also does not establish that this build reproduces Unsloth's calibration process or measured quality. This is an independent community build.

Security calibration

The reviewed calibration material covers public cybersecurity text, source code, tool exchanges, general text and selected investigation evidence.

The security suffix describes the calibration focus. Importance-matrix collection informs quantization; this workflow performs no additional gradient-based fine-tuning or distillation. Improved vulnerability discovery or tool reliability has not been measured for this quantization.

The build scripts use a 4,096-token calibration context. That setting is separate from the context configured when serving the model. The exact corpus and completed imatrix run used for the final release are not yet pinned in this card.

MTP and local inference

MTP enables speculative decoding in a runtime that supports the model's prediction head. Keeping the head in the GGUF and enabling speculative decoding are separate requirements.

Current llama.cpp documentation names the MTP mode draft-mtp. Check llama-server --help for the options supported by your installed build.

Example for PowerShell, using a compatible CUDA-enabled llama.cpp build:

$ModelPath = ".\Feyospace-v1.1-Abliterated-security-MTP-UD-Q4_K_S.gguf"

.\llama-server.exe --model $ModelPath --alias feyospace-security --ctx-size 32768 --n-gpu-layers 99 --parallel 1 --jinja --spec-type draft-mtp --spec-draft-n-max 3 --host 127.0.0.1 --port 8080

Adjust the path to the downloaded filename. The example requests GPU offload and a 32,768-token serving context; it is an illustrative configuration, not a measured memory-fit or throughput result. Tune context and offload for the available hardware.

The local API base URL is http://127.0.0.1:8080/v1, with model alias feyospace-security. Use the model's supplied chat template and the runtime's tool-calling interface when connecting an agent. See the llama.cpp server documentation for template and API options.

To establish a baseline without speculative decoding, use --spec-type none and omit --spec-draft-n-max 3. Measure MTP on the same prompts and serving settings before drawing a speed comparison.

Context and memory

The original Feyospace model card reports a configured maximum context of 262,144 tokens. That is an upstream configuration value, not a tested context guarantee for this GGUF.

Runtime memory includes model weights, context caches, compute buffers and MTP-related allocations. This card makes no fixed VRAM requirement, maximum practical context or tokens-per-second claim.

Evaluation and build status

No benchmark results for this exact quantization are available in the material used to prepare this card.

The original Feyospace and OBLITERATUS repositories publish upstream evaluations. Those results describe their respective checkpoints, precisions and evaluation setups; they are not measurements of this GGUF.

The build scripts document intended checks and processing steps. Completed build logs and the final model binary were not available for this card. The following remain unverified:

• Final GGUF size, SHA-256 and complete tensor inventory, including MTP tensor types.
• Exact source revision, calibration corpus hash and completed imatrix provenance.
• Successful loading, tool-call parsing and MTP execution on a named runtime build.
• Quality retention, long-context behavior and MTP speedup.

Intended use and limitations

Intended uses include authorized vulnerability research, source and binary analysis, controlled reproduction experiments, CTF work, harness development and evidence-based report drafting.

Abliteration is inherited from OBLITERATUS and is intended to reduce refusals. It does not establish technical correctness or guarantee any particular response behavior. Validate generated code, proposed vulnerabilities and material report claims against actual artifacts and execution results.

When using the model as an agent, enforce scope, tool permissions and isolation in the surrounding system. Model instructions are not an access-control boundary.

Attribution and license

• Feyospace Team / Vera Praxis: original cybersecurity model and research.
• OBLITERATUS: the source checkpoint's cybersecurity-focused abliteration.
• Unsloth: the mixed-precision layout reference.
• llama.cpp contributors: conversion, quantization and inference tooling.

The Apache-2.0 metadata follows the OBLITERATUS source repository. Preserve applicable upstream license and attribution files when distributing the weights.