IQ4_NL is a non-linear quantization format. It performs sub-optimally on Apple's Metal backend due to the lack of native support for non-linear dequantization kernels. If you are running on an Apple Silicon Mac with GPU offloading via Metal, you will likely experience:
Slower inference compared to linear quants of similar size (e.g., Q4_K_M)
No speed benefit from the ARM weight repacking that IQ4_NL supports on CPU
If you're on Apple Metal, use the Q4_K_M quant from this repo instead. For higher precision options, see Jongsim's repo.
A reasoning-enhanced, abliterated version of Qwen3.5-35B-A3B (35B total / 3B active parameters, Mixture of Experts). This model was built in two stages: first, censorship removal via directional ablation using Heretic, then supervised fine-tuning on high-quality Chain-of-Thought reasoning traces distilled from Claude 4.6 Opus.
The model produces structured reasoning within <think>...</think> tags before delivering final responses. All weights are in bf16 precision.
The primary objective is to inject high-density structured reasoning capability from Claude 4.6 Opus while preserving the uncensored nature of the abliterated base model. Through SFT on curated reasoning distillation data, the model learns to decompose complex problems into sequential steps within a dedicated thinking block before generating the final answer.
The base model was processed with Heretic v1.2.0, an automated censorship removal tool that applies directional ablation optimized via Bayesian hyperparameter search (Optuna TPE).
Two techniques were combined:
SOMA (Self-Organizing Map Abliteration): Uses a 4x4 SOM to discover multiple refusal directions in activation space, then ablates the top-k directions simultaneously.
MPOA (Magnitude-Preserving Orthogonal Ablation): Projects out the refusal direction while preserving the original weight magnitude via row normalization with low-rank correction (rank 4).
Abliteration Configuration
Parameter
Value
Method
SOMA + MPOA
Orthogonalize Direction
true
Row Normalization
full
Full Normalization LoRA Rank
4
Winsorization Quantile
0.95
SOM Grid
4 x 4 (16 neurons)
SOM Iterations
10,000
SOM Learning Rate
0.01
SOM Sigma
0.5
SOM k (directions)
4
Optimization Trials
200 (60 startup)
Selected Trial
Trial 84 / 200
Good Prompts
mlabonne/harmless_alpaca (train[:400])
Bad Prompts
mlabonne/harmful_behaviors (train[:400])
Quantization
none (bf16)
Abliteration Results
Metric
Original
Abliterated
KL Divergence
0 (reference)
0.0638
Refusals (out of 100)
91
6
93.4% refusal reduction with minimal distribution shift (KL = 0.0638).
Inject structured Chain-of-Thought reasoning patterns from Claude 4.6 Opus into the abliterated model. The training enforces a strict output format where the model generates internal reasoning within <think> blocks before producing the final response.
Training Strategy
Framework: Unsloth 2026.3.3 + TRL SFTTrainer
Method: LoRA (Low-Rank Adaptation) applied to both attention and MoE expert layers
Loss Computation: train_on_responses_only — loss is calculated exclusively on assistant responses (both thinking trace and final answer), not on user prompts
Instruction boundary: <|im_start|>user\n
Response boundary: <|im_start|>assistant\n<think>
Chat Template: Qwen ChatML format (<|im_start|> / <|im_end|>)
LoRA Configuration
Parameter
Value
PEFT Method
LoRA
Rank (r)
16
Alpha
32 (= 2 x rank)
Dropout
0.0
Bias
none
Target Modules (Attention)
q_proj, k_proj, v_proj, o_proj
Target Modules (FFN)
gate_proj, up_proj, down_proj
Target Modules (MoE)
gate_up_proj
Gradient Checkpointing
unsloth mode
Training Hyperparameters
Parameter
Value
Max Sequence Length
2,048
Per-Device Batch Size
1
Gradient Accumulation Steps
8
Effective Batch Size
8
Number of Epochs
5
Total Training Steps
1,995
Learning Rate
2e-4
LR Scheduler
Linear decay
Warmup Steps
5
Optimizer
AdamW 8-bit
Weight Decay
0.001
Precision
bf16
Seed
3407
Total FLOPs
3.56 x 10^18
Datasets
Three publicly available reasoning distillation datasets were combined, shuffled (seed=42), and used for training:
Curated reasoning samples in both conversation and instruction format, designed for step-by-step problem solving diversity.
Total
~3,191
Combined after filtering empty/invalid rows.
Training Loss
Epoch
Avg Loss
Steps
1
0.4299
79 - 399
2
0.3729
400 - 798
3
0.3359
799 - 1197
4
0.3059
1198 - 1596
5
0.2958
1597 - 1995
Training loss decreased monotonically from 0.4299 to 0.2958 across 5 epochs, indicating stable convergence without overfitting signs at the loss level.
Checkpoint Selection
The best checkpoint was selected based on GSM8K accuracy (50 samples). All checkpoints were evaluated in isolated subprocesses to prevent GPU memory leaks from Unsloth's model patching.
Checkpoint
Epoch
GSM8K Accuracy
checkpoint-1200
3.0
8.0% (4/50)
checkpoint-1400
3.5
10.0% (5/50)
checkpoint-1596
4.0
10.0% (5/50)
checkpoint-1995
5.0
12.0% (6/50)
checkpoint-1995 (epoch 5) was selected and merged into bf16 for the final release.
Note: GSM8K measures narrow arithmetic reasoning and does not fully reflect the model's broader reasoning capabilities (code generation, logical analysis, multi-step planning) which are the primary targets of the distillation training.
Hardware and Environment
Component
Value
Hardware
NVIDIA DGX Spark
GPU
NVIDIA GB10 (128GB unified memory)
Compute Capability
sm121
Architecture
aarch64
CUDA
13.0
PyTorch
2.9.1a0
Transformers
5.2.0
Unsloth
2026.3.3
TRL
0.24.0
PEFT
0.18.1
Datasets
4.3.0
Tokenizers
0.22.2
DGX Spark-Specific Notes
Flash Attention and Memory-Efficient Attention (cutlass) are disabled due to sm121 incompatibility (supported: sm80-sm100). Only Math SDP is used.
flash_attn package is fully removed to prevent FATAL errors on sm121.
torch.compile / TorchInductor is disabled due to Triton ptxas compatibility issues.
The entire model (35B parameters) fits in a single GPU's 128GB unified memory without quantization.
Usage
This model uses the standard Qwen3.5 chat template. It operates in thinking mode by default.
Inference Example
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model_name ="Jongsim/Qwen3.5-35B-A3B-heretic-Reasoning"4model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")5tokenizer = AutoTokenizer.from_pretrained(model_name)67messages =[8{"role":"user","content":"Explain the difference between TCP and UDP, and when you would choose one over the other."}9]1011text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)12inputs = tokenizer(text, return_tensors="pt").to(model.device)13outputs = model.generate(**inputs, max_new_tokens=8192, temperature=0.7, top_p=0.8, top_k=20)14response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)15print(response)
Recommended Sampling Parameters
Mode
temperature
top_p
top_k
presence_penalty
Thinking (general)
1.0
0.95
20
1.5
Thinking (coding)
0.6
0.95
20
0.0
Non-thinking (general)
0.7
0.8
20
1.5
Example of Learned Reasoning Format
The model produces output in the following structure:
<think>
Let me analyze this problem step by step.
1. First, I need to identify the core question being asked.
2. Then, I'll consider the relevant constraints and conditions.
3. Next, I'll work through the logic systematically.
4. Finally, I'll verify my reasoning for consistency.
[detailed reasoning follows...]
</think>
[final answer here]
This structured thinking pattern, distilled from Claude 4.6 Opus interactions, reduces redundant cognitive loops while preserving deep analytical capacity.
Limitations
Hallucination Risk: As an autoregressive language model, the model may generate plausible-sounding but factually incorrect statements, particularly regarding real-world events or obscure technical details.
GSM8K Performance: The model scores 12% on GSM8K (50 samples). This is expected because the training data emphasizes broad reasoning patterns (code, logic, planning) rather than arithmetic drill. For pure math benchmarks, consider models specifically trained on mathematical datasets.
Abliteration Residual: 6 out of 100 harmful prompts still trigger refusal. The abliteration is not exhaustive.
Context Length Trade-off: While the architecture supports 262K tokens natively, the SFT was performed with max_seq_length=2048. Very long reasoning chains beyond the training distribution may degrade in quality.
MoE Inference Overhead: Despite having only 3B active parameters per token, the full 35B model must be loaded into memory. Minimum ~65GB VRAM/RAM required for bf16.
Acknowledgements
Qwen Team for the Qwen3.5-35B-A3B architecture and pretrained weights
Heretic (p-e-w) for the automated directional ablation framework
Unsloth AI for efficient LoRA fine-tuning of large MoE models
