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.

Qwen3.5-0.8B Abliterated: MNN Format (mix-mask export)

MNN conversion of huihui-ai/Huihui-Qwen3.5-0.8B-abliterated (vision included) for on-device inference. All credit for the abliteration goes to huihui-ai.

• Hybrid attention: 18 LinearAttention + 6 full-attention layers of 24: self-describing metadata (attention_type:"mix", layer_nums:24, sliding_window:4), stacked mask graph [2,1,1,seq,kv]
• Quantization: 4-bit, block 64, HQQ (matches the taobao-mnn 0.8B bundle parameters)
• Tokenizer: tokenizer.mtok (new format)
• Exporter: MNN-TokForge tokforge/rebase-3.5.0 @ 21c339c2fd (upstream #4336 attention-type detection)
• Requires an MNN runtime with the mix-mask gen_attention_mask path (MNN-TokForge ≥ tokforge/rebase-3.5.0); legacy single-mask runtimes are NOT compatible
• Validated: multi-turn coherence (3 cycles/session, CPU), LinearAttention snapshot/restore (Y=Z), self-describing hybrid metadata verified against the exported graph

File checksums (md5)

3af7de552fb9cea05e67f615d96efb53  llm.mnn
10fe7c624159f62c69d93a0e836fc80f  llm.mnn.weight
88bbc9851fc9e2aae364eaf59698326a  llm_config.json
3c183b4e10bef80909cb6998fe5e4fa6  tokenizer.mtok
06b782392bdb4ca19a1045c3709fdbb6  visual.mnn
e48e427a48dc363278ecaacb7810db6a  visual.mnn.weight

Credits

• Huihui-Qwen3.5-0.8B-abliterated by huihui-ai (Hugging Face). Used under the Apache-2.0 license.
• Qwen3.5-0.8B 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).