license: apache-2.0
language:
- en
base_model: - prithivMLmods/Qwen3.6-35B-A3B-abliterated-MAX
pipeline_tag: text-generation
tags: - abliterated
- qwen3.6
- APEX
- GGUF
- unfiltered
- compressed
- nano
- uncensored
- reasoning
- max
- prithivMLmods
Qwen3.6-35B-A3B-abliterated-MAX | APEX i-nano (2.72 BPW)
This model was quantized using apex-quant with the i-nano profile and an importance matrix calibrated on a diverse code/math/reasoning dataset.
Quantization Details
| Property | Value |
|---|---|
| Base Model | prithivMLmods/Qwen3.6-35B-A3B-abliterated-MAX |
| Quantizer | mudler/apex-quant |
| Profile | i-nano (importance-matrix calibrated) |
| BPW | 2.72 |
| File Size | ~11 GB |
| Layers | 40 |
| Calibration Data | tomngdev/imatrix-calibration-data |
What is APEX Quantization?
APEX applies a per-layer, per-tensor quantization gradient _ higher precision on edge layers (first and last ~5), aggressive quantization on the middle layers, with separate handling for routed experts, shared experts, attention weights, and SSM weights. The i-nano variant uses importance matrix calibration to enable very low-bit formats (IQ2_S, IQ2_XXS) on middle-layer expert weights while preserving output quality.
Usage
Run with any recent llama.cpp build, no custom fork or patches required:
# CLI
./llama-cli -m Qwen3.6-35B-A3B-abliterated-MAX-APEX-i-nano.gguf -p "Your prompt here"
# Server
./llama-server -m Qwen3.6-35B-A3B-abliterated-MAX-APEX-i-nano.gguf--host 0.0.0.0 --port 8080
Files
| File | Description |
|---|---|
Qwen3.6-35B-A3B-abliterated-MAX-APEX-i-nano.gguf |
The quantized model (~11 GB) |
imatrix.dat |
Importance matrix used for calibration |