GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
w3joe
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
w3joe
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
All four models were evaluated on the same NVIDIA GB10 hardware in bfloat16 with greedy decoding (temperature = 0.0), identical prompt splits, and identical lm-evaluation-harness protocols across 1,438 held-out items per model (raw JSON receipts).
| Metric / Benchmark | Items | Qwen/Qwen3-8B (Base) | Qwen3-8B-abliterated (Ours) | Huihui-Qwen3-8B-v2 | Heretic 2.0 (ARA) |
|---|---|---|---|---|---|
| Harmful Refusal Rate ↓ | 300 | 99.33% (298/300) | 0.67% (2/300) | 2.67% (8/300) | 2.67% (8/300) |
| Benign Over-Refusal ↓ | 200 | 1.00% (2/200) | 1.50% (3/200) | 1.00% (2/200) | 1.50% (3/200) |
| Benign Next-Token KL ↓ | 200 | 0.0000 | 0.0919 | 0.1041 | 0.2799 |
| GSM8K (Strict Match) ↑ | 128 | 68.75% | 69.53% | 67.19% | 60.16% |
| MMLU-Pro (14 Subjects) ↑ | 112 | 38.39% | 42.86% | 39.29% | 37.50% |
| MMLU (57 Subjects) ↑ | 570 | 62.11% | 61.93% | 61.58% | 46.84% |
| IFEval (Prompt Strict) ↑ | 128 | 61.72% | 58.59% | 58.59% | 53.91% |
| Average Capability ↑ | 938 | 57.74% | 58.23% (+0.49 pp) | 56.66% (-1.08 pp) | 49.60% (-8.14 pp) |
| Benchmark | Items | Ours vs. Base Qwen3-8B | Ours vs. Huihui-v2 | Ours vs. Heretic 2.0 |
|---|---|---|---|---|
| GSM8K | 128 | +0.78 pp [-1.56, +3.91] | +2.34 pp [-1.56, +6.25] | +9.38 pp [+3.12, +15.62] |
| MMLU-Pro | 112 | +4.46 pp [-0.89, +9.82] | +3.57 pp [-0.89, +8.93] | +5.36 pp [-0.89, +11.61] |
| MMLU | 570 | -0.18 pp [-1.58, +1.23] | +0.35 pp [-1.23, +1.93] | +15.09 pp [+11.93, +18.07] |
| IFEval | 128 | -3.12 pp [-7.03, +0.00] | 0.00 pp [-3.91, +3.91] | +4.69 pp [-0.80, +10.94] |
| File Path | Format | Size | Recommended Use |
|---|---|---|---|
model-00001-of-00004.safetensors .. 00004 | BF16 Safetensors | 16.4 GB | Transformers, vLLM, SGLang, LoRA fine-tuning |
gguf/Qwen3-8B-abliterated-Q8_0.gguf | Q8_0 GGUF | 8.7 GB | Near-lossless local inference (Ollama, LM Studio, llama.cpp) |
gguf/Qwen3-8B-abliterated-Q4_K_M.gguf | Q4_K_M GGUF | 5.0 GB | Fast 4-bit inference for 8 GB+ VRAM & Apple Silicon |
gguf/Qwen3-8B-abliterated-f16.gguf | F16 GGUF | 16.4 GB | Unquantized GGUF reference |
bash
# Run directly with llama.cppllama-cli -hf w3joe/Qwen3-8B-abliterated:Q8_0 --jinja -cnv -p "You are a helpful assistant."To import into Ollama using the included gguf/Modelfile:
bash
huggingface-cli download w3joe/Qwen3-8B-abliterated --include "gguf/*" --local-dir ./Qwen3-8B-abliteratedcd Qwen3-8B-abliterated/ggufollama create qwen3-8b-abliterated -f Modelfileollama run qwen3-8b-abliteratedpython
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "w3joe/Qwen3-8B-abliterated"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto",) messages = [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Explain how heap spraying works in memory corruption exploits."},]prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, # Set True to enable <think> reasoning chains)inputs = tokenizer([prompt], return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))bash
vllm serve w3joe/Qwen3-8B-abliterated --dtype bfloat16 --max-model-len 32768Standard abliteration subtracts the raw difference of means (refusal_mean - benign_mean) across all layers, which shifts weight row norms and damages directions responsible for math and instruction following.
Qwen3-8B-abliterated applies three refinements during weight editing:
q = 0.99): Per-prompt residual stream activations are clamped at the 99th percentile in FP32 before averaging so outlier coordinates do not skew the extracted direction.rank = 4): At each layer, the refusal vector is orthogonalized not only against the mean harmless activation, but also against the top-4 principal components (SVD) of the benign activation matrix—preventing the edit from encroaching on reasoning and instruction-following subspaces.L7–L35, strength = 0.85) + Exact Row-Norm Preservation: Early syntax/embedding layers (L0–L6, where benign and refusal activations are >99.7% collinear) are left untouched, and every modified row in self_attn.o_proj and mlp.down_proj is rescaled in FP32 to preserve its exact original Euclidean norm before saving to bfloat16.Qwen/Qwen3-8B).# Run directly with llama.cppllama-cli -hf w3joe/Qwen3-8B-abliterated:Q8_0 --jinja -cnv -p "You are a helpful assistant."huggingface-cli download w3joe/Qwen3-8B-abliterated --include "gguf/*" --local-dir ./Qwen3-8B-abliteratedcd Qwen3-8B-abliterated/ggufollama create qwen3-8b-abliterated -f Modelfileollama run qwen3-8b-abliteratedimport torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "w3joe/Qwen3-8B-abliterated"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto",) messages = [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Explain how heap spraying works in memory corruption exploits."},]prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, # Set True to enable <think> reasoning chains)inputs = tokenizer([prompt], return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))vllm serve w3joe/Qwen3-8B-abliterated --dtype bfloat16 --max-model-len 32768