Published by TokForge for on-device inference.

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.

Qwen3.5-9B Abliterated: MNN Format

This is an MNN-converted version of huihui-ai/Huihui-Qwen3.5-9B-abliterated for on-device mobile inference.

All credit for the abliteration work goes to huihui-ai. We only performed the MNN conversion and quantization for mobile deployment.

What is this?

• Base model: Qwen/Qwen3.5-9B by Alibaba
• Abliteration by: huihui-ai: removes refusal behavior via orthogonal projection (FailSpy technique)
• MNN conversion by: darkmaniac7: 4-bit quantization (block size 128) for mobile GPU/CPU inference
• Purpose: On-device roleplay, creative fiction, and mature content without refusal. Richer writing and deeper character interactions than the 4B variant.

Model Details

Usage

This model is designed for TokForge, an offline Android AI chat app. It can also be used with any MNN-compatible runtime.

TokForge (Android)

Models → Recommended → Roleplay → "Qwen3.5 9B Uncensored" → Download

Manual

Download all files and load with MNN's llm_demo or the MNN Transformer API.

Limitations and Intended Use

• Intended for TokForge / MNN mobile inference and local roleplay-style use.
• Qwen3.5 LinearAttention models route differently from standard Qwen3 targets and may prefer CPU on some phones.
• Large-model mobile performance depends heavily on device memory pressure and backend routing.
• This repo is a mobile runtime/export artifact, not a standard Transformers release.

Files

Credits

• Huihui-Qwen3.5-9B-abliterated by huihui-ai (Hugging Face). Used under the Apache-2.0 license.
• Qwen3.5-9B by the Qwen team (Alibaba Cloud) (Hugging Face). Used under the Apache-2.0 license.
• The MTP head is exported from huihui-ai/Huihui-Qwen3.5-9B-abliterated, which carries the MTP tensors of Qwen/Qwen3.5-9B unchanged (Apache-2.0), by TokForge with llmexport from TokForge's fork of MNN.
• Converted with llmexport from MNN by Alibaba (Apache-2.0).

Community

• Website: tokforge.ai
• Discord: Join the Discord

License

Apache 2.0 (inherited from Qwen3.5)

MTP head (optional add-on)

The same head files as darkmaniac7/Qwen3.5-9B-MNN, byte for byte, for the Qwen3.5 9B Abliterated trunk in this repo (llm.mnn.weight sha256 d07c7a30cefa0aa24b4de7261f79a13436908323b884c05c531321e0a9a40d20). Exported from huihui-ai/Huihui-Qwen3.5-9B-abliterated @ 05b9e7c9b978ba29bdb8f50a49c30e4b91183339, which still carries Qwen/Qwen3.5-9B's MTP tensors byte for byte with MNN-TokForge llmexport.py --path <HF checkpoint> --export mnn --mtp --mtp_only --quant_bit 4 --quant_block 64 --mtp_vocab_file vocab_v32k.json (32k draft vocabulary, mtp-v32k.mnn.vocab, int32 token ids). Load with speculative_type: "mtp", mtp_model: "mtp-v32k.mnn", draft_predict_length: 1. The trunk checks every drafted token, so the head never adds a token the trunk did not pick. Host acceptance with the trunk in this repo (MNN CPU route, greedy, depth 1, 96 tokens; share of drafted tokens the trunk accepted): code 94%, advice 68%, chat 73%, RAG 83%. TokForge checks on each phone whether it makes replies faster before it uses it.