license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.6-27B-abliterated
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: mnn
language:
- en
- zh
tags: - tokforge
- on-device
- mnn
- chat
- qwen
- qwen3.6
- qwen3_5
- image-text-to-text
- vlm
- multimodal
- abliterated
Huihui Qwen3.6 27B abliterated (MNN)
huihui-ai's abliterated Qwen3.6 27B as a 4-bit MNN bundle, with its vision encoder, for 24 GB phones running TokForge.
Published by TokForge for on-device inference.
| What it is | Vision chat model (text + images; 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 | 24 GB RAM for practical use; 16 GB phones load it but time out on longer prompts |
| Download | 18.9 GB (11 files) |
| Base model | huihui-ai/Huihui-Qwen3.6-27B-abliterated |
| License | Apache-2.0 |
TokForge
- Website: https://tokforge.ai
- Discord: https://discord.gg/Acv3CBtfVm
- Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
- iOS TestFlight: https://testflight.apple.com/join/jnufjzRr
Runs on-device in the TokForge app.
Use it in TokForge
- Install TokForge from Google Play, or join the iOS TestFlight.
- On Android, open Models > Browse, search for
Huihui-Qwen3.6-27B-abliterated-MNNand download it. - Load it and start a chat. The model runs on the phone; once it is downloaded, chatting needs no connection.
Use it with upstream MNN llm_demo
pip install -U huggingface_hub
hf download darkmaniac7/Huihui-Qwen3.6-27B-abliterated-MNN --local-dir Huihui-Qwen3.6-27B-abliterated-MNN
git 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/Huihui-Qwen3.6-27B-abliterated-MNN/config.json prompt.txt
Collections: TokForge Qwen Models (MNN), TokForge Vision Models.
Files
| File | Size | What it is |
|---|---|---|
config.json |
387 B | MNN runtime settings |
embeddings_bf16.bin |
2.54 GB | Token embeddings (bf16) |
export_args.json |
1.12 KB | Export settings used for the conversion |
llm.mnn |
8.87 MB | Model graph |
llm.mnn.json |
22.4 MB | Model graph as JSON (reference copy) |
llm.mnn.weight |
16.0 GB | Model weights |
llm_config.json |
8.66 KB | Model settings and chat template |
tokenizer.txt |
6.47 MB | Tokenizer |
tokforge_benchmarks.json |
20.1 KB | Raw summary of TokForge's device checks (April and May 2026) |
visual.mnn |
825 KB | Vision encoder graph (Qwen3.6 vision sidecar) |
visual.mnn.weight |
261 MB | Vision encoder weights |
About the model
MNN-format 4-bit TokForge export of huihui-ai/Huihui-Qwen3.6-27B-abliterated, packaged for Android on-device inference with TokForge.
| Source checkpoint | huihui-ai/Huihui-Qwen3.6-27B-abliterated |
| Base model | Qwen/Qwen3.6-27B |
| Source snapshot SHA | 27502c8717fd5a2f8c0c77188c10c243fd4f672e |
| Architecture | qwen3_5, 64 layers (48 of them linear attention), hidden size 5120, 24 attention heads, 4 KV heads, hybrid linear + full attention, MRoPE, DeepStack vision sidecar |
Vision
The bundle includes the vision encoder (visual.mnn, visual.mnn.weight), and llm_config.json marks the model as
a vision model (is_visual). TokForge's device checks covered text generation only.
Device notes
In TokForge's device checks (April 30 and May 1, 2026), the 24 GB phones (RedMagic, Redmi) completed every prompt and
are practical for short local text generation with this export. The 16 GB phones can load the model, but longer
paragraph or code prompts are not practical in this CPU MNN configuration: OnePlus and Lenovo completed short and code
prompts but timed out on paragraph generation; S26 and Pixel completed only arithmetic and a short sentence before
timing out on longer prompts.
How it was made
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 |
Runtime defaults in config.json: CPU backend, 4 threads, low precision, low memory, separated BF16 embeddings.
Limitations
- This is a safety-reduced / abliterated model; deploy with appropriate product policy, user controls, and local-law awareness.
- Large and slow on phones: TokForge runs it on the CPU by default (the GPU route is opt-in), and it is practical only on 24 GB phones.
- Image input is included in the bundle but was not part of TokForge's device checks.
- This is a runtime bundle, not a standard Transformers training checkpoint.
License
Apache-2.0, inherited from the upstream Qwen and huihui-ai model cards. See the
Apache License 2.0.
Credits
- Huihui-Qwen3.6-27B-abliterated by huihui-ai (Hugging Face). Used under the Apache-2.0 license.
- Qwen3.6-27B by the Qwen team (Alibaba Cloud) (Hugging Face). Used under the Apache-2.0 license.
- Converted with llmexport from MNN by Alibaba (Apache-2.0).