base_model:
- xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
license: apache-2.0
language: - en
pipeline_tag: image-text-to-text
library_name: transformers
tags: - mtp
- text-generation-inference
- llama-cpp
- transformers
- unsloth
- qwen3_5
- reasoning
- distillation
- deepseek
- sft
- rl
- gspo
- math
- stem
- tool-use
- function-calling
- heretic
- uncensored
- decensored
- abliterated
- speculative-decoding
- nvidia-blackwell
- rtx-pro-6000
- SM120
- GB202
- B200
- B300
- FP4
- blackwell
- Qwen3.5
- GPTQ
datasets: - HuggingFaceH4/ultrachat_200k
DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-NVFP4-GPTQ
DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-NVFP4-GPTQ is a 4-bit NVFP4 GPTQ quantization of DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic, a Heretic v2.0.0.dev0 decensored model based on Jackrong's Qwen3.5-9B reasoning fine-tune distilled from DeepSeek-V4-Pro. The original decensoring reduced refusals from 100/100 to 3/100 at a KL divergence of 0.0422. GPTQ calibration helps preserve model accuracy while reducing memory requirements and improving inference efficiency on NVFP4-capable hardware. Reported original-model benchmarks include 94.50% on GSM8K and 90.53% on a MMLU-Pro STEM sample. These results are not independently verified for this quantized build. The base weights are licensed under Apache 2.0. Experimental outputs require independent verification.
Quantization Recipe
The model is quantized using the GPTQModifier with the NVFP4 quantization scheme.
default_stage:
default_modifiers:
GPTQModifier:
targets: [Linear]
ignore: [lm_head, "re:.*visual.*"]
scheme: NVFP4
weight_observer: imatrix_mse
bypass_divisibility_checks: false
requires_calibration_data: true
block_size: 128
dampening_frac: 0.01
actorder: static
batched_quantization: auto
Model Files
| File Name | File Type | File Size | File Link | Description |
|---|---|---|---|---|
| chat_template.jinja | Jinja Template | 8.74 kB | Link | Chat template for formatting conversations and prompts. |
| config.json | JSON | 9.58 kB | Link | Core model architecture and configuration parameters. |
| generation_config.json | JSON | 142 Bytes | Link | Default text generation parameters (e.g., temperature, top_p). |
| model-00001-of-00002.safetensors | Safetensors | 4.99 GB | Link | Model weight shard 1/2. Required to load the model. |
| model-00002-of-00002.safetensors | Safetensors | 3.89 GB | Link | Model weight shard 2/2. Required to load the model. |
| model.safetensors.index.json | JSON | 153 kB | Link | Index file mapping model weights to their respective shards. |
| model_mtp.safetensors | Safetensors | 487 MB | Link | Multi-Token Prediction (MTP) auxiliary model weights, used to speed up generation. |
| processor_config.json | JSON | 1.33 kB | Link | Configuration for the multimodal processor (if applicable). |
| recipe.yaml | YAML | 357 Bytes | Link | Configuration file detailing the quantization or training recipe (e.g., GPTQ parameters). |
| tokenizer.json | JSON | 20 MB | Link | Complete tokenizer vocabulary and encoding rules. |
| tokenizer_config.json | JSON | 1.53 kB | Link | Tokenizer-specific configuration parameters. |
License and Attribution
This model is based on and/or incorporates the following open-source projects and models:
- Qwen3.5-9B (Base): https://huggingface.co/Qwen/Qwen3.5-9B
- DeepSeek-V4-Pro-Qwen3.5-9B: https://huggingface.co/Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B
- DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic: https://huggingface.co/xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
- llm-compressor: https://github.com/vllm-project/llm-compressor
This model is released under the Apache License 2.0.