license: apache-2.0
base_model: sensenova/SenseNova-Loop-9B
language:
- fr
- en
pipeline_tag: text-generation
tags: - gguf
- llama.cpp
- q4_k_m
SenseNova-Loop-9B Uncensored — Q4_K_M GGUF
| Metric | Result |
|---|---|
| Base-model fidelity | ≈97% |
| Refusals | ≈3 / 100 |
| Factual accuracy | 30 / 30 (tested in direct and reasoning modes) |
| Context window | 512K tokens (base model) |
| GGUF size | 5.6 GB |
A 9B reasoning model in a single 5.6 GB GGUF. This uncensored edition of SenseNova-Loop-9B runs locally, supports English and French, and keeps the original chat template for direct answers and deeper reasoning.
Tested with the GuideboardLabs llama.cpp fork using HIP on an AMD Radeon RX 9070 XT.
Release notes
This checkpoint was created by applying targeted weight ablation to the release checkpoint, then quantizing it to Q4_K_M. The edit was designed to reduce refusals while preserving bilingual text generation and reasoning. The original chat template was carried over.
Run it locally
Find SenseNova-Loop-9B-Uncensored-Q4_K_M.gguf under Files and versions. Load it with a runtime that supports sensenova_loop, such as the GuideboardLabs llama.cpp fork. Practical context length depends on available memory and runtime settings.
Evaluation and limitations
The factual scores come from a small internal test set. Reasoning can take a long time, and characters from other languages can occasionally appear in French responses. The weight edit can change response behavior and reliability.
License
Apache 2.0. The base model was published by SenseTime / OpenSenseNova.