TokForge Acceleration Pack Qwen3.5 0.8B draft (MNN)
A deprecated Qwen3.5 0.8B draft model for speculative decoding, kept on the Hub for reproducibility; TokForge no longer uses it.
Published by TokForge for on-device inference.
| What it is | Draft model for speculative decoding (deprecated) |
| Format | MNN bundle, 4-bit weights (block 64) |
| Runs on | Android phones (CPU), MNN runtime; not used by the current TokForge app |
| Phone class | No catalog class; needs more free RAM than its 470 MB of weights |
| Download | 484 MB (7 files) |
| Base model | Qwen/Qwen3.5-0.8B |
| 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.
- TokForge does not offer or download this repo. For Qwen3 chat models, open Settings > Speed & engine > Models & Modules and download Acceleration Pack: Qwen3-0.6B 40k (TokForge-AccelerationPack-Draft) instead.
- Qwen3.5 chat models need no separate draft: TokForge's Qwen3.5 MNN downloads carry an optional MTP head (the model's own multi-token prediction layer), for example darkmaniac7/Qwen3.5-4B-MNN.
This repo is for TokForge / MNN users who specifically want to reproduce the older Qwen3.5-0.8B draft path.
Collections: TokForge Draft Models.
Files
| File | Size | What it is |
|---|---|---|
config.json |
342 B | MNN runtime settings |
export_args.json |
1.10 KB | Export settings used for the conversion |
llm.mnn |
2.15 MB | Model graph |
llm.mnn.json |
5.34 MB | Model graph as JSON (reference copy) |
llm.mnn.weight |
470 MB | Model weights |
llm_config.json |
8.69 KB | Model settings and chat template |
tokenizer.txt |
6.47 MB | Tokenizer |
Why it is deprecated
Deprecated: this
Qwen3.5-0.8Bdraft bundle is preserved for reproducibility, but the newerQwen3-0.6BTokForge draft line is the practical default.
- Cross-family draft pairing does not attach: Qwen3.5 and Qwen3 use different vocabularies (248,320 vs 151,936), so this 0.8B draft cannot pair with Qwen3 targets, and the Qwen3-0.6B drafts cannot pair with Qwen3.5 targets.
- Qwen3.5 models use linear attention. TokForge keeps speculative decoding with a separate draft model off for them and uses their MTP head instead.
Use instead
- For Qwen3 chat models: TokForge-AccelerationPack-Draft (in the app: Acceleration Pack: Qwen3-0.6B 40k).
- For Qwen3.5 chat models: the optional MTP head in TokForge's Qwen3.5 MNN downloads, for example darkmaniac7/Qwen3.5-4B-MNN and darkmaniac7/Qwen3.5-9B-MNN.
How it was made
Exported with MNN's llmexport from huihui-ai's abliterated Qwen3.5 0.8B (huihui-ai/Huihui-Qwen3.5-0.8B-abliterated) as a text-only model (the vision part switched off,
is_visual=false), with 4-bit weights in blocks of 64 and tied embeddings. export_args.json has the full export
settings.
Limitations
- Deprecated for normal use.
- Cross-family drafting is not possible (vocabulary mismatch), and TokForge does not use separate draft models with Qwen3.5 targets.
- Keep using this only if you need reproducibility for old experiments.
- This is an MNN runtime bundle, not a standard Transformers checkpoint.
License
Apache License 2.0 (text), the license of Qwen3.5 0.8B.
Credits
- Qwen3.5-0.8B by the Qwen team (Alibaba Cloud) (Hugging Face). Used under the Apache-2.0 license.
- Huihui-Qwen3.5-0.8B-abliterated by huihui-ai (Hugging Face). Used under the Apache-2.0 license.
- Converted with llmexport from MNN by Alibaba (Apache-2.0).
- Downloads last month
- 73