license: apache-2.0
language:
- en
pipeline_tag: image-text-to-text
tags: - chat
- qwen
- qwen3.6
- qwen3_5
- image-text-to-text
- vlm
- multimodal
- abliterated
- uncensored
- mnn
- tokforge
base_model: - huihui-ai/Huihui-Qwen3.6-27B-abliterated
- Qwen/Qwen3.6-27B
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.
Huihui-Qwen3.6-27B-abliterated-MNN
MNN-format 4-bit TokForge export of huihui-ai/Huihui-Qwen3.6-27B-abliterated, packaged for Android on-device inference with TokForge.
What this is
- Source checkpoint:
huihui-ai/Huihui-Qwen3.6-27B-abliterated - Base model:
Qwen/Qwen3.6-27B - Source snapshot SHA:
27502c8717fd5a2f8c0c77188c10c243fd4f672e - Architecture:
qwen3_5, 64 layers, hidden size 5120, 24 attention heads, 4 KV heads, hybrid linear + full attention, MRoPE, DeepStack vision sidecar. - Target runtime: TokForge Android MNN runtime. Join the TokForge community on Discord.
The bundle includes the visual MNN sidecar, but the validation below is text-generation only through TokForge's /test-prompt endpoint.
Bundle contents
config.json- TokForge/MNN runtime defaults.llm_config.json- model geometry and tokenizer/chat settings.llm.mnn/llm.mnn.weight- quantized LLM graph and external weights.embeddings_bf16.bin- separated BF16 embedding table.tokenizer.txt- tokenizer exported by the MNN pipeline.visual.mnn/visual.mnn.weight- Qwen3.6 vision sidecar.export_args.json- conversion settings.llm.mnn.json- exported graph metadata.tokforge_benchmarks.json- raw benchmark summary from the device validation run.
Quantization scheme
| 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.
TokForge device validation
Validated on April 30 and May 1, 2026 over wireless ADB with TokForge's authenticated /test-prompt control endpoint. The prompts were intentionally independent rather than sequential chat turns: arithmetic, short sentence, paragraph generation, and a small Python function.
Generation config for every prompt:
| Setting | Value |
|---|---|
| Backend | mnn |
| Context | 8192 |
| Threads | 4 |
| Precision | low |
thinking_enabled |
false |
temperature |
0.0 |
top_p |
1.0 |
top_k |
1 |
seed |
42 |
| Path reported by TokForge | full_history |
| QNN/NPU | not requested, not attempted, CPU effective |
Summary
| Device | Model | SoC | RAM | Load time | Avg decode tok/s | Weighted decode tok/s | Avg prefill | Completed | Coherent completed | Timeouts |
|---|---|---|---|---|---|---|---|---|---|---|
| RedMagic | NX809J |
SM8850 |
24 GB | ~31-33 s | 2.49 | 2.25 | 12.06 s | 4/4 | 4/4 | - |
| Redmi | 2407FRK8EC |
MT6989 |
24 GB | 55 s | 2.63 | 2.33 | 25.66 s | 4/4 | 4/4 | - |
| OnePlus | PLC110 |
MT6991 |
16 GB | 51.80 s | 0.28 | 0.24 | 25.19 s | 3/4 | 3/3 | paragraph |
| Lenovo | TB520FU |
SM8650 |
16 GB | 62.90 s | 0.30 | 0.27 | 20.07 s | 3/4 | 3/3 | paragraph |
| S26 | SM-S948U1 |
SM8850 |
16 GB | 29.47 s | 0.25 | 0.22 | 17.93 s | 2/4 | 2/2 | paragraph, code |
| Pixel | Pixel 9 Pro XL |
Tensor G4 |
16 GB | 76.55 s | 0.12 | 0.11 | 33.08 s | 2/4 | 2/2 | paragraph, code |
Weighted decode tok/s is computed as total generated decode tokens divided by total decode wall time. For devices with timeouts, decode averages are computed across completed prompts only.
The 24 GB devices completed the full four-prompt matrix and are practical for short local text generation with this export. The 16 GB devices can load the model, but longer paragraph/code prompts are not practical in this CPU MNN configuration: OnePlus and Lenovo completed short/code prompts but timed out on paragraph generation; S26 and Pixel completed only arithmetic and short sentence before timing out on longer prompts.
Per-prompt decode results
| Device | Prompt class | Prompt tokens | Decode tokens | Decode tok/s | Result |
|---|---|---|---|---|---|
| RedMagic | arithmetic | 73 | 3 | 3.07 | 12 |
| RedMagic | short sentence | 72 | 14 | 2.58 | coherent |
| RedMagic | paragraph | 83 | 122 | 2.24 | coherent 5-sentence paragraph |
| RedMagic | Python code | 87 | 20 | 2.06 | valid clamp function |
| Redmi | arithmetic | 92 | 3 | 3.42 | 12 |
| Redmi | short sentence | 91 | 14 | 2.43 | coherent |
| Redmi | paragraph | 102 | 103 | 2.29 | coherent 5-sentence paragraph |
| Redmi | Python code | 106 | 20 | 2.37 | valid clamp function |
| OnePlus | arithmetic | 92 | 3 | 0.36 | 12 |
| OnePlus | short sentence | 91 | 14 | 0.25 | coherent |
| OnePlus | paragraph | - | - | timeout | no final response captured; server history showed partial decode in the 56-60 token range |
| OnePlus | Python code | 106 | 20 | 0.23 | valid clamp function |
| Lenovo | arithmetic | 92 | 3 | 0.37 | 12 |
| Lenovo | short sentence | 91 | 14 | 0.27 | coherent |
| Lenovo | paragraph | - | - | timeout | HTTP 504 after roughly 300 s; server generated 71 tokens before timeout |
| Lenovo | Python code | 106 | 20 | 0.26 | valid clamp function |
| S26 | arithmetic | 92 | 3 | 0.29 | 12 |
| S26 | short sentence | 91 | 14 | 0.20 | coherent |
| S26 | paragraph | - | - | timeout | HTTP 504 / no final response |
| S26 | Python code | - | - | timeout | HTTP 504 / no final response |
| Pixel | arithmetic | 92 | 3 | 0.14 | 12 |
| Pixel | short sentence | 91 | 14 | 0.10 | coherent |
| Pixel | paragraph | - | - | timeout | HTTP 504 / no final response |
| Pixel | Python code | - | - | timeout | HTTP 504 / no final response |
The paragraph and code prompts were added to reduce benchmark bias from tiny outputs. Completed code prompts produced the expected Python function:
def clamp(value, low, high):
return max(low, min(value, high))
Download
pip install -U huggingface_hub
hf download darkmaniac7/Huihui-Qwen3.6-27B-abliterated-MNN --local-dir Huihui-Qwen3.6-27B-abliterated-MNN
Usage with upstream MNN llm_demo
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
Host smoke validation with the TokForge MNN build loaded successfully, detected 48 linear-attention state layers out of 64 total layers, and emitted the expected one-token OK response for a no-thinking prompt.
Attribution
- Original base model:
Qwen/Qwen3.6-27B - Abliterated checkpoint:
huihui-ai/Huihui-Qwen3.6-27B-abliterated - Export tooling: alibaba/MNN LLM export, TokForge MNN fork and Android runtime.
- Conversion, packaging, and Android validation: TokForge /
darkmaniac7.
License and safety
Apache-2.0, inherited from the upstream Qwen and huihui-ai model cards. This is a safety-reduced / uncensored model; deploy with appropriate product policy, user controls, and local-law awareness.