DreamFast

qwen3-8b-heretic

README

License: apache-2.0

Model Details

Base Model: Qwen/Qwen3-8B
Abliteration Method: Heretic v1.2.0 (git master, commit 19cdf7e)
Trials: 3000
Trial Selected: Trial 2681
Refusals: 13/100 (vs 100/100 original)
KL Divergence: 0.0838 (minimal model damage)

Files

HuggingFace Format (for transformers, llama.cpp conversion)

markdown
model.safetensors (~16 GB)
config.json
tokenizer.json
tokenizer_config.json
generation_config.json
chat_template.jinja

ComfyUI Format (for Klein 9B text encoder)

markdown
comfyui/qwen3-8b-heretic.safetensors              # bf16, 16GB
comfyui/qwen3-8b-heretic_fp8_e4m3fn.safetensors   # fp8 per-tensor, 8.8GB
comfyui/qwen3-8b-heretic_int8.safetensors          # int8 ConvRot row-wise, 8.8GB
comfyui/qwen3-8b-heretic_int4.safetensors          # int4 W4A4 ConvRot, 5.6GB
comfyui/qwen3-8b-heretic_nvfp4.safetensors         # nvfp4, 6.0GB
comfyui/qwen3-8b-heretic_mxfp8.safetensors         # mxfp8, 9.0GB

Quality: All quantized variants use SVD-guided learned rounding (AdaRound via convert-to-quant), which optimizes each weight's rounding direction to minimize output reconstruction error, noticeably higher fidelity than naive round-to-nearest quantization.

GGUF Format (for llama.cpp and ComfyUI-GGUF)

Table with columns: Quant, Size, Notes
Quant	Size	Notes
F16	16GB	Lossless reference
Q8_0	8.2GB	Excellent quality
Q6_K	6.3GB	Very good quality
Q5_K_M	5.5GB	Good quality
Q5_K_S	5.4GB	Slightly smaller Q5
Q4_K_M

Quantization Format Notes

All variants load natively in ComfyUI 0.30.0+ (no plugins) via the comfy_quant metadata embedded in each file.

Table with columns: Format, Size, Bits, Notes
Format	Size	Bits	Notes
FP8 (E4M3)	8.8GB	8	Per-tensor scaled, learned rounding; Ada/Hopper+
INT8 (ConvRot row-wise)	8.8GB	8	Hadamard-rotated, per-row; broad GPU support
MXFP8	9.0GB	8	Microscaling FP8 (E8M0 block scales); Blackwell
INT4 (W4A4 ConvRot)	5.6GB	4

NVFP4/MXFP8 inference is fastest on Blackwell (RTX 5090/5080, SM100+), but ComfyUI also supports software dequantization on older GPUs (tested on RTX 4090). INT8 and INT4 both use Hadamard rotation (ConvRot); INT4 W4A4 uses ComfyUI's convrot_w4a4 path.

Usage

With ComfyUI (Klein 9B)

Download a ComfyUI format file:
- FP8 (recommended): comfyui/qwen3-8b-heretic_fp8_e4m3fn.safetensors (8.8GB)
- INT4 (smallest): comfyui/qwen3-8b-heretic_int4.safetensors (5.6GB)
- NVFP4: comfyui/qwen3-8b-heretic_nvfp4.safetensors (6.0GB)
- INT8: comfyui/qwen3-8b-heretic_int8.safetensors (8.8GB)
- MXFP8: comfyui/qwen3-8b-heretic_mxfp8.safetensors (9.0GB)
- bf16 (full precision): comfyui/qwen3-8b-heretic.safetensors (16GB)
Place in

With Transformers

python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "DreamFast/qwen3-8b-heretic",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("DreamFast/qwen3-8b-heretic")

prompt = "Describe a dramatic sunset over a cyberpunk city"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With llama.cpp

bash
llama-server -m qwen3-8b-heretic-Q4_K_M.gguf

Abliteration Process

Created using Heretic v1.2.0 (git master) with 3000 optimization trials:

markdown
? Which trial do you want to use?
  [Trial 2732] Refusals: 10/100, KL divergence: 0.1001
> [Trial 2681] Refusals: 13/100, KL divergence: 0.0838  <-- selected
  [Trial 2337] Refusals: 18/100, KL divergence: 0.0643
  [Trial 2419] Refusals: 19/100, KL divergence: 0.0600
  [Trial 2195] Refusals: 21/100, KL divergence: 0.0534
  ...

Trial 2681 was selected for its balance of low refusals (13/100) and reasonable KL divergence (0.0838), indicating minimal model damage while achieving 87% of previously-refused prompts now working.

Limitations

This model inherits all limitations of the base Qwen3-8B model
Abliteration reduces but does not completely eliminate refusals (13/100 remain)
NVFP4 quantization works best on Blackwell GPUs (RTX 5090/5080) with native FP4 tensor cores, but also works on older GPUs via software dequantization

License

This model is released under the Apache 2.0 License, following the base Qwen3-8B model license.

Acknowledgments

Qwen for the Qwen3-8B model
Heretic by p-e-w for the abliteration tool
Black Forest Labs for Klein 9B
llama.cpp for GGUF conversion

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

DreamFast

Model Tree

Base

Qwen/Qwen3-8B

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Base Model: Qwen/Qwen3-8B
Abliteration Method: Heretic v1.2.0 (git master, commit 19cdf7e)
Trials: 3000
Trial Selected: Trial 2681
Refusals: 13/100 (vs 100/100 original)
KL Divergence: 0.0838 (minimal model damage)

Files

HuggingFace Format (for transformers, llama.cpp conversion)

markdown
model.safetensors (~16 GB)
config.json
tokenizer.json
tokenizer_config.json
generation_config.json
chat_template.jinja

ComfyUI Format (for Klein 9B text encoder)

markdown
comfyui/qwen3-8b-heretic.safetensors              # bf16, 16GB
comfyui/qwen3-8b-heretic_fp8_e4m3fn.safetensors   # fp8 per-tensor, 8.8GB
comfyui/qwen3-8b-heretic_int8.safetensors          # int8 ConvRot row-wise, 8.8GB
comfyui/qwen3-8b-heretic_int4.safetensors          # int4 W4A4 ConvRot, 5.6GB
comfyui/qwen3-8b-heretic_nvfp4.safetensors         # nvfp4, 6.0GB
comfyui/qwen3-8b-heretic_mxfp8.safetensors         # mxfp8, 9.0GB

Quality: All quantized variants use SVD-guided learned rounding (AdaRound via convert-to-quant), which optimizes each weight's rounding direction to minimize output reconstruction error, noticeably higher fidelity than naive round-to-nearest quantization.

GGUF Format (for llama.cpp and ComfyUI-GGUF)

Table with columns: Quant, Size, Notes
Quant	Size	Notes
F16	16GB	Lossless reference
Q8_0	8.2GB	Excellent quality
Q6_K	6.3GB	Very good quality
Q5_K_M	5.5GB	Good quality
Q5_K_S	5.4GB	Slightly smaller Q5
Q4_K_M

Quantization Format Notes

All variants load natively in ComfyUI 0.30.0+ (no plugins) via the comfy_quant metadata embedded in each file.

Table with columns: Format, Size, Bits, Notes
Format	Size	Bits	Notes
FP8 (E4M3)	8.8GB	8	Per-tensor scaled, learned rounding; Ada/Hopper+
INT8 (ConvRot row-wise)	8.8GB	8	Hadamard-rotated, per-row; broad GPU support
MXFP8	9.0GB	8	Microscaling FP8 (E8M0 block scales); Blackwell
INT4 (W4A4 ConvRot)	5.6GB	4

Usage

With ComfyUI (Klein 9B)

Download a ComfyUI format file:
- FP8 (recommended): comfyui/qwen3-8b-heretic_fp8_e4m3fn.safetensors (8.8GB)
- INT4 (smallest): comfyui/qwen3-8b-heretic_int4.safetensors (5.6GB)
- NVFP4: comfyui/qwen3-8b-heretic_nvfp4.safetensors (6.0GB)
- INT8: comfyui/qwen3-8b-heretic_int8.safetensors (8.8GB)
- MXFP8: comfyui/qwen3-8b-heretic_mxfp8.safetensors (9.0GB)
- bf16 (full precision): comfyui/qwen3-8b-heretic.safetensors (16GB)
Place in

With Transformers

python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "DreamFast/qwen3-8b-heretic",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("DreamFast/qwen3-8b-heretic")

prompt = "Describe a dramatic sunset over a cyberpunk city"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With llama.cpp

bash
llama-server -m qwen3-8b-heretic-Q4_K_M.gguf

Abliteration Process

Created using Heretic v1.2.0 (git master) with 3000 optimization trials:

markdown
? Which trial do you want to use?
  [Trial 2732] Refusals: 10/100, KL divergence: 0.1001
> [Trial 2681] Refusals: 13/100, KL divergence: 0.0838  <-- selected
  [Trial 2337] Refusals: 18/100, KL divergence: 0.0643
  [Trial 2419] Refusals: 19/100, KL divergence: 0.0600
  [Trial 2195] Refusals: 21/100, KL divergence: 0.0534
  ...

Trial 2681 was selected for its balance of low refusals (13/100) and reasonable KL divergence (0.0838), indicating minimal model damage while achieving 87% of previously-refused prompts now working.

Limitations

This model inherits all limitations of the base Qwen3-8B model
Abliteration reduces but does not completely eliminate refusals (13/100 remain)
NVFP4 quantization works best on Blackwell GPUs (RTX 5090/5080) with native FP4 tensor cores, but also works on older GPUs via software dequantization

License

This model is released under the Apache 2.0 License, following the base Qwen3-8B model license.

Acknowledgments

Qwen for the Qwen3-8B model
Heretic by p-e-w for the abliteration tool
Black Forest Labs for Klein 9B
llama.cpp for GGUF conversion