Downloads · 30 days
150
14% of all-time downloads
darkmaniac7/Qwen3.6-35B-A3B-abliterated-MNN
Qwen3.6-35B-A3B-abliterated-MNN is a image-text-to-text model from darkmaniac7. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for mnn. The card lists the license as apache-2.0.
Youssofal's Heretic-abliterated Qwen3.6 35B-A3B, a multimodal mixture-of-experts model, as a 4-bit MNN bundle for phones with 16 GB RAM or more in TokForge.
Downloads · 30 days
150
14% of all-time downloads
All-time downloads
1.1K
Public
Repo size
22.8 GB
Likes
2
Public
Click a slice to open those files.
.weight21.6 GB · 95%
From the Hugging Face model README
Youssofal's Heretic-abliterated Qwen3.6 35B-A3B, a multimodal mixture-of-experts model, as a 4-bit MNN bundle for phones with 16 GB RAM or more in TokForge.
Published by TokForge for on-device inference.
| What it is | Vision chat model (text + images; MoE, 35B total, 3B active; abliterated) |
| Format | MNN bundle, HQQ 4-bit weights (block 64), with vision encoder |
| Runs on | Android phones (CPU); also MNN's llm_demo on a computer |
| Phone class | 16 GB RAM or more (TokForge catalog; opt-in below 24 GB) |
| Download | 22.8 GB (11 files) |
| Base model | Youssofal/Qwen3.6-35B-A3B-Abliterated-Heretic-BF16 |
| License | Apache-2.0 |
Runs on-device in the TokForge app.
llm_demogit clone https://github.com/alibaba/MNN.git
cd MNN && mkdir build && cd build
cmake .. -DMNN_LOW_MEMORY=true -DMNN_CPU_WEIGHT_DEQUANT_GEMM=true \
-DMNN_BUILD_LLM=true -DMNN_SUPPORT_TRANSFORMER_FUSE=true
make -j
./llm_demo /path/to/Qwen3.6-35B-A3B-abliterated-MNN/config.json prompt.txt
Collections: TokForge Qwen Models (MNN), TokForge Vision Models.
| File | Size | What it is |
|---|---|---|
config.json | 382 B | MNN runtime settings |
embeddings_bf16.bin | 1.02 GB | Token embeddings (bf16) |
export_args.json | 1.12 KB | Export settings used for the conversion |
llm.mnn | 54.3 MB | Model graph |
llm.mnn.json | 119 MB | Model graph as JSON (reference copy) |
llm.mnn.weight | 21.3 GB | Model weights |
llm_config.json | 8.64 KB | Model settings and chat template |
tokenizer.mtok | 7.32 MB | Tokenizer (MNN binary format; the one config.json uses) |
tokenizer.txt | 2.96 MB | Tokenizer (text format) |
visual.mnn | 548 KB | Vision encoder graph (DeepStack) |
visual.mnn.weight | 254 MB | Vision encoder weights |
MNN-format 4-bit quantization of the Heretic-abliterated Qwen3.6-35B-A3B multimodal MoE, packaged for TokForge.
| Source (upstream abliteration) | Youssofal/Qwen3.6-35B-A3B-Abliterated-Heretic-BF16 |
| Upstream of that | Qwen/Qwen3.6-35B-A3B |
| Architecture | qwen3_5_moe (40 layers, 256 experts, 8 active per token, hybrid linear + full attention, DeepStack vision) |
| Parameters | 35B total, about 3B active per token (4-bit quantized) |
Abliteration methodology: Heretic input-side split-MoE transfer (MPOA/SOMA-style). Upstream reports 1/25 refusal rate on the official 25-prompt harmful-behaviors check (vs. 22/25 for the base model) with KL-divergence 0.0107 vs. base.
TokForge keeps memory use low by paging inactive experts from flash storage.
This bundle ships with visual.mnn + visual.mnn.weight, the DeepStack vision transformer for image input.
llm_config.json has is_visual: true with the full vision config block (image_mean, image_norm, image_size: 420,
vision_start, vision_end, image_pad, num_grid_per_side: 48, has_deepstack: true).
The vision tower in Qwen3.6-35B-A3B is architecturally identical across the base Qwen release and the Heretic-abliterated variant (the abliteration changed only the text-side MoE layers).
Attention-stack fields in llm_config.json: attention_type: mix, sliding_window: 4, layer_nums: 40.
Exported with MNN's llmexport (settings in export_args.json):
| Flag | Value |
|---|---|
--quant_bit | 4 |
--quant_block | 64 |
--lm_quant_bit | 4 |
--lm_quant_block | 64 |
--embed_bit | 16 |
--hqq | enabled |
--seperate_embed | enabled |
The vision files were exported in a second pass with the same tools. Export pipeline: llmexport from TokForge's fork of MNN (upstream exporter: alibaba/MNN llmexport).
The tokenizer ships in two forms: tokenizer.mtok (MNN's binary tokenizer, the one config.json names) and
tokenizer.txt. Running the bundle needs an MNN build that reads .mtok tokenizers and DeepStack vision models;
TokForge's engine does both.