license: apache-2.0
base_model: virtualdmns/SmolLM3-3B-Heretic-Uncensored
library_name: gguf
tags:
- gguf
- quantized
- smollm3
- heretic
- uncensored
SmolLM3-3B-Heretic-Uncensored-GGUF
GGUF builds of virtualdmns/SmolLM3-3B-Heretic-Uncensored, the abliterated (refusal-removed) SmolLM3-3B produced with Heretic. Converted from the verified safetensors weights with llama.cpp at commit 8e5a58df87ab59dfceb5b961ae26691e39f30c24 (convert_hf_to_gguf.py via an F16 intermediate, then llama-quantize). The model's chat template is embedded in each file, including the /think and /no_think modes. For the full method, benchmark, and evaluation details, see the main repo.
Files
| File | Size (bytes) | Notes |
|---|---|---|
SmolLM3-3B-Heretic-Uncensored-Q8_0.gguf |
3,275,574,720 | Near lossless. Recommended if your RAM allows. |
SmolLM3-3B-Heretic-Uncensored-Q4_K_M.gguf |
1,915,305,408 | Smaller and faster, with some quality loss from quantization. |
Usage
With llama.cpp:
# One-shot generation
./llama-cli -m SmolLM3-3B-Heretic-Uncensored-Q8_0.gguf \
-p "Explain why the sky is blue in three sentences." -n 256
# Interactive conversation
./llama-cli -m SmolLM3-3B-Heretic-Uncensored-Q8_0.gguf -i
Any GGUF runner works: llama.cpp, LM Studio, or Ollama (via a Modelfile import of the .gguf file).
Verification
- Both files were load tested in llama.cpp (
llama-cli): they load, use the embedded chat template, and generate coherently. Test generations included a sky explanation and a pin tumbler lock mechanism walkthrough, answered plainly with no refusals, in both quantizations. - The weights they were converted from were independently re-evaluated after export: refusals 3/100 and KL divergence 0.0684, against a stock SmolLM3-3B baseline of 82/100 refusals.
Known limitations
- On severe harmful prompts under greedy decoding, the underlying model often falls into repetitive token loops instead of answering coherently. Benign and refusal-adjacent generation was unaffected in all spot checks.
Licenses
- Base model weights: Apache-2.0 (HuggingFaceTB/SmolLM3-3B). These builds are released under the same terms.
- Heretic, the tool used to produce the underlying model, is AGPL-3.0 software. That license applies to the Heretic code, not to these weights.