Downloads · 30 days
24
32% of all-time downloads
Academica/Academica-Qwen3-8B-3Bit-Heretic
Academica-Qwen3-8B-3Bit-Heretic is a text generation model from Academica. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
Qwen3-8B with Heretic refusal-direction ablation and Q2K quantization.
Downloads · 30 days
24
32% of all-time downloads
All-time downloads
75
Public
Repo size
3.3 GB
Likes
0
Public
Click a slice to open those files.
.gguf3.3 GB · 100%
From the Hugging Face model README
Qwen3-8B with Heretic refusal-direction ablation and Q2_K quantization.
Academica produced the ablated model and all three published quantizations. This 3-bit release balances memory use, throughput, and generation quality.
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3-8B |
| Modification | Heretic ablation after 20 Optuna trials |
| Quantization | Q2_K |
| File size | 3.1 GB |
| Measured throughput | 58 tokens/second |
| Measured perplexity | 1.86 |
| Refusal evaluation | 18 refusals from 100 prompts |
| Evaluation hardware | NVIDIA DGX Spark, GB10 Blackwell, 128 GB unified memory |
The measurements describe the published artifact on the stated hardware. Results can differ with prompts, runtimes, and sampling settings.
qwen3-8b-heretic-q2_k.gguf: Q2_K model weights.from huggingface_hub import hf_hub_download
model_path = hf_hub_download(
repo_id="Academica/Academica-Qwen3-8B-3Bit-Heretic",
filename="qwen3-8b-heretic-q2_k.gguf",
)
llama-server -m qwen3-8b-heretic-q2_k.gguf --host 0.0.0.0 --port 8080 -ngl 99
Q2_K compression can reduce accuracy relative to larger quantizations. Validate the model against the intended task and runtime.
The ablation reduces refusal behavior. It does not establish factual accuracy, safety, or suitability for regulated decisions.