Huihui-Nex-N2-mini
Abliterated Agentic Vision MoE — APEX Quantized GGUF
⚡ Thinking Mode Requires Nex's Patched llama.cppNex-N2-mini's original chat template uses complex vision processing macros that stock llama.cpp's Jinja parser cannot handle correctly. This causes thinking tags to not be injected, breaking thinking mode output.
The official fix: Use Nex's patched llama.cpp, which works with the unmodified GGUF and unmodified template. Once Nex's upstream patch is merged into stock llama.cpp, this workaround will no longer be needed.
⚠️ Do NOT modify chat_template.jinja. The model was trained strictly on the current template — editing the tags deviates from the training-time format and may degrade output quality. See discussion #3.
This is an abliterated (uncensored) version of Nex-N2-mini created by huihui-ai. Safety filtering has been significantly reduced.
- May generate sensitive, controversial, or inappropriate content
- Not suitable for public settings, underage users, or production use
- Users are solely responsible for compliance with local laws and ethical standards
- Recommended for research, testing, or controlled environments only
Original model: huihui-ai/Huihui-Nex-N2-mini-abliterated
💡 What is APEX?These GGUF files are quantized using APEX, a MoE-aware mixed-precision quantization technique that outperforms standard quantization methods while being significantly smaller.
APEX beats Q8_0 perplexity at half the size — and even beats F16.
APEX classifies every tensor by its role — routed expert, shared expert, or attention — and applies a layer-wise precision gradient, giving the most sensitive edge layers higher precision and compressing the redundant middle layers more aggressively.
📦 Available Files| File | Size | BPW | Note |
|---|---|---|---|
Huihui-Nex-N2-mini-abliterated.BF16.gguf | 64.6 GB | 16.0 | Full precision reference |
Huihui-Nex-N2-mini-abliterated-APEX-Quality.gguf | 21.3 GB | 5.23 | Highest quality, best accuracy |
Huihui-Nex-N2-mini-abliterated-APEX-Balanced.gguf | 23.6 GB | 5.85 | Best all-rounder, recommended |
Huihui-Nex-N2-mini-abliterated-APEX-Compact.gguf | 15.4 GB | 3.81 | Best quality/size ratio, 16GB VRAM |
mmproj-Nex-N2-mini.F16.gguf | 858 MB | - | Vision projector (required for image/video) |
chat_template.jinja | 7.9 KB | - | Original unmodified chat template |
| Architecture | Qwen3.5 MoE (GatedDeltaNet + Full Attention) + Vision Encoder |
| Parameters | 35B total, 3B active per token |
| Experts | 256 routed experts, 8 active per token |
| Layers | 40 layers (30 linear_attn + 10 full_attn) |
| Context | 262,144 tokens |
| Vision | Image support (mmproj 858MB) |
| Thinking | Qwen3-style think tags — requires Nex's patched llama.cpp |
| Modification | Abliterated by huihui-ai (safety filters removed) |
Download Nex's patched llama.cpp
Binaries: nex-agi/llama.cpp | Docker: ghcr.io/nex-agi/llama.cpp:server-cuda-nex-b9596-fix-b9598-8c0d5c9
./llama-server \ -m Huihui-Nex-N2-mini-abliterated-APEX-Quality.gguf \ -ngl 99 -ncmoe 19 -c 32768 \ --host 0.0.0.0 --port 8081
Replace Huihui-Nex-N2-mini-abliterated-APEX-Quality.gguf with your preferred quantization tier (Quality / Balanced / Compact). Add --mmproj mmproj-Nex-N2-mini.F16.gguf for vision. Recommended sampling: temperature 0.7, top_p 0.95, top_k 40, min_p 0.
| Benchmark | Score | Category |
|---|---|---|
| BrowseComp | 74.1 | Agent |
| SWE-Bench Verified | 74.4 | Coding |
| Terminal-Bench 2.1 | 60.7 | Coding |
| GPQA Diamond | 82.6 | Reasoning |
| IFEval | 89.1 | Instruction |
From the original Nex-N2-mini model card (BF16, full precision). Abliteration does not significantly change benchmark scores.
Links
- APEX Quantization: https://github.com/mudler/apex-quant
- Abliterated Model: https://huggingface.co/huihui-ai/Huihui-Nex-N2-mini-abliterated
- Original Model: https://huggingface.co/nex-agi/Nex-N2-mini
- Nex's Patched llama.cpp (required for thinking mode): https://github.com/nex-agi/llama.cpp/releases/tag/nex-b9596-fix-b9599-9cd1771
- Original Model Discussion: https://huggingface.co/nex-agi/Nex-N2-Pro/discussions/3
- Abliteration Technique: https://github.com/Sumandora/remove-refusals-with-transformers
- Qwen3.5-35B-A3B: https://huggingface.co/Qwen/Qwen3.5-35B-A3B