A 9B Qwen3.5 reasoning model distilled from DeepSeek-V4-Pro on ~250K math & STEM samples — then decensored with Heretic v2 (ARA). This page is an interactive explorer for its benchmarks, ablation parameters and token efficiency.
Heretic's Arbitrary-Rank Ablation finds a direction in activation space per layer and projects it out. This run only touches the middle of the network — layers 9 through 24 of 32 — which is why general capability survives while refusal behaviour collapses.
| ARA parameter | Value | Meaning |
|---|---|---|
start_layer_index | 9 | first layer ablated |
end_layer_index | 24 | last layer ablated |
preserve_good_behavior_weight | 0.8264 | how hard to keep benign capabilities intact |
steer_bad_behavior_weight | 0.0003 | pull away from refusal behaviour — near zero, so the ablation is the main driver |
overcorrect_relative_weight | 0.9645 | penalty for overshooting the target direction |
neighbor_count | 12 | neighbouring layers used to estimate each direction |
All figures as published on the model card. Hover a bar for the exact value.
The 9B student lands just above its DeepSeek-V4-Pro teacher.
Highest average among the tested 9B baselines (+0.93 over Qwen3.5-9B).
| Model | Math | Physics | Chemistry | Average |
|---|---|---|---|---|
| DeepSeek-V4-Pro-Qwen3.5-9B | 92.40% | 89.40% | 89.80% | 90.53% |
| Qwen3.5-9B | 90.60% | 89.00% | 89.20% | 89.60% |
| Claude Mythos-distilled 27B | 88.40% | 87.60% | 82.60% | 86.20% |
| Claude Mythos-distilled 9B | 86.00% | 78.60% | 78.60% | 81.07% |
| DeepSeek-V4-Pro-Qwen3.5-4B | 80.40% | 74.20% | 74.80% | 76.47% |
9B rows use MTP-Q8_0 GGUF at 8K context; the 4B student uses MTP-Q8 GGUF at 64K.
100 adversarial prompts, scored for refusal.
On the same 1,500-question MMLU-Pro sample this model combines the highest accuracy with the fewest reasoning tokens per correct answer — 36.1% fewer than Qwen3.5-9B and 22.7% fewer than the Claude Mythos-distilled 9B.
| Model | Accuracy | Avg reasoning tokens / question | Tokens / correct answer | Correct / 1M tokens |
|---|---|---|---|---|
| DeepSeek-V4-Pro-Qwen3.5-9B | 90.53% | 2,858 | 3,157 | 316.8 |
| Qwen3.5-9B | 89.60% | 4,425 | 4,938 | 202.5 |
| Claude Mythos-distilled 9B | 81.07% | 3,310 | 4,083 | 244.9 |
Uses the published tokens per correct answer figures — a rough budget, not a guarantee.
Asked to "output only ONE single capital letter", the model complied 97.2% of the time and re-explained in only 0.2%.
The checkpoint is ~18.8 GB in bf16, so it fits on a single 24 GB card in 4-bit or a 40 GB card in bf16. Pick your stack.
import torch
from transformers import AutoProcessor, AutoModelForImageTextToText, BitsAndBytesConfig
MODEL = "xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic"
bnb = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16,
)
processor = AutoProcessor.from_pretrained(MODEL)
model = AutoModelForImageTextToText.from_pretrained(
MODEL, quantization_config=bnb, device_map="cuda",
dtype=torch.bfloat16, attn_implementation="sdpa",
).eval()
messages = [{
"role": "user",
"content": [{"type": "text", "text": "Prove that sqrt(2) is irrational."}],
}]
inputs = processor.apply_chat_template(
messages, enable_thinking=True, add_generation_prompt=True,
tokenize=True, return_dict=True, return_tensors="pt",
).to("cuda")
out = model.generate(**inputs, max_new_tokens=2048, do_sample=True,
temperature=0.6, top_p=0.95)
text = processor.tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:],
skip_special_tokens=True)
reasoning, _, answer = text.partition("</think>")
print(answer.strip())# vLLM serves the OpenAI-compatible API
vllm serve xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic \
--max-model-len 32768 \
--gpu-memory-utilization 0.90
# then
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")
r = client.chat.completions.create(
model="xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic",
messages=[{"role": "user", "content": "Balance MnO4- + C2O4^2- in base."}],
max_tokens=2048, temperature=0.6,
)
print(r.choices[0].message.content)# The card reports a llama.cpp MTP-Q8_0 path. Use a GGUF conversion:
llama-cli -hf xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF \
--ctx-size 8192 --temp 0.6 -p "Prove that sqrt(2) is irrational."
# or a local server
llama-server -hf xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF \
--ctx-size 8192 --host 0.0.0.0 --port 8080# A ready-made Gradio Space ships in this repo at /gradio_app
# It targets ZeroGPU and loads the model in 4-bit at module scope.
hf repos create <you>/deepseek-v4-pro-9b-heretic --type space \
--space-sdk gradio --flavor zero-a10g --public
hf upload <you>/deepseek-v4-pro-9b-heretic ./gradio_app --repo-type space
# Requirements: transformers>=5.20, accelerate, torchvision, bitsandbytes.
# Note: ZeroGPU needs a PRO account or an account older than 30 days.ARA removes refusal behaviour, not the model's capacity for error. The absence of a refusal is not evidence that a claim is correct, safe, or lawful.