Qwen3.5 27B Claude 4.6 Opus abliterated (MNN)
huihui-ai's abliteration of Jackrong's Claude 4.6 Opus reasoning distill of Qwen3.5 27B, as a 4-bit MNN bundle for phones with a lot of memory in TokForge.
Published by TokForge for on-device inference.
| What it is | Reasoning chat model |
| Format | MNN bundle, 4-bit weights (block 128), bf16 embeddings, text only |
| Runs on | Android phones (CPU) |
| Phone class | 24 GB RAM or more recommended |
| Download | 17.0 GB (8 files) |
| Base model | huihui-ai/Huihui-Qwen3.5-27B-Claude-4.6-Opus-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
Qwen3.5-27B-Claude-4.6-Opus-abliterated-MNNand download it. - Load it and start a chat. The model runs on the phone; once it is downloaded, chatting needs no connection.
Collections: TokForge Qwen Models (MNN).
Files
| File | Size | What it is |
|---|---|---|
config.json |
210 B | MNN runtime settings |
embeddings_bf16.bin |
2.54 GB | Token embeddings (bf16, required) |
export_args.json |
1.11 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 |
14.4 GB | Model weights |
llm_config.json |
4.43 KB | Model settings and chat template |
tokenizer.txt |
2.96 MB | Tokenizer |
embeddings_bf16.bin is required: the model reads its token embeddings from it.
About the model
Pre-converted Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated (huihui-ai's abliteration of Jackrong's Claude 4.6 Opus reasoning distill of Qwen3.5-27B) in MNN format for on-device inference.
Jackrong trained Qwen3.5 27B on reasoning outputs from Claude 4.6 Opus; huihui-ai then removed its refusal behavior (abliteration), so it declines fewer requests. It thinks step by step before it answers.
| Architecture | Qwen3.5 (hybrid LinearAttention + standard attention), text only |
| Parameters | 27B (4-bit quantized) |
| Quantization | W4A16 (4-bit weights, 16-bit activations), block size 128 |
| Attention | 48 LinearAttention layers (gated_delta_rule) + 16 standard attention layers |
| RAM | 24GB+ device recommended |
| Backend | CPU (TokForge's default for Qwen3.5 LinearAttention models) |
Limitations
- Intended for TokForge / MNN power users with very high-memory devices.
Qwen3.5hybrid LinearAttention models are usually better on CPU than OpenCL in current TokForge routing.- This is a runtime bundle for on-device inference, not a standard Transformers training repo.
- Large-model mobile performance depends heavily on device RAM, swap behavior, and OEM background-killer policies.
- Abliterated: refusal behavior removed. Use responsibly.
License
Apache-2.0, as Qwen3.5 27B, Jackrong's distill and huihui-ai's abliteration.
Credits
- Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated by huihui-ai (Hugging Face). Used under the Apache-2.0 license. Our MNN files were converted from this model.
- Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled by Jackrong (Hugging Face). Used under the Apache-2.0 license. Reasoning-distilled from Claude 4.6 Opus outputs; not an Anthropic release.
- Qwen3.5-27B 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).