Qwen3.5-2B-MNN (mirror with MTP head)
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.
This repo mirrors taobao-mnn/Qwen3.5-2B-MNN at commit 35781816d7b6a9dcb273a6765ac9563401951c3c byte for byte
(every file below has the same sha256 as the original; the MNN conversion and all credit for it belong to the
taobao-mnn team, the model to the Qwen team) and adds an optional multi-token-prediction (MTP) head next to the
unchanged trunk, used by TokForge on the MNN CPU route.
Trunk files (identical to taobao-mnn/Qwen3.5-2B-MNN @ 35781816d7)
| file | bytes | sha256 |
|---|---|---|
config.json |
652 | 92853033efe602f95efca3e1c05cd8b108f973c8beed417843a9671f8147ed8d |
export_args.json |
1,040 | a5b3a7d2c45e53c887e539259696d5fcec793637c9a3e9fbc7cd0dd008af6a87 |
llm.mnn |
2,148,136 | 23df98f8b341b277365e0bbca025c1d192939e3d32d7f79776352c6f32e77960 |
llm.mnn.json |
5,344,018 | 7131ff4f1a441add1039d371815ad94652ae6801f579591cb2f9ad90b5954025 |
llm.mnn.weight |
1,176,647,702 | c93f71a2dbecf9328782bd38861656d8faa82e95e7f99607350074768a482054 |
llm_config.json |
8,692 | a88234b36c2af0eff8e5c89667011badf71c15e30459eb0e21030a8f3f9ed240 |
tokenizer.txt |
6,465,727 | 7e75de1f279a10b65bd9dc1a5207205cb8993823861c4c42bbbd74e48e1c23a4 |
visual.mnn |
488,096 | 88fc40a7b676e90eb2cb86d854db15cb90b9eb1f34087ab0f48c5e43572c8dac |
visual.mnn.weight |
195,587,264 | 8f90e106f5b9ae9a939faed240305cfdd5c6740ae91d3fc418a990bee0cce36b |
MTP head (optional add-on)
Exported from Qwen/Qwen3.5-2B @ 15852e8c16360a2fea060d615a32b45270f8a8fc 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 98%, advice 56%, chat 58%, RAG 76%. TokForge checks on each phone whether it makes replies faster before it uses it.
| file | bytes | sha256 |
|---|---|---|
mtp-v32k.mnn.vocab |
132,160 | 80af0c499240994c092bb3e79fe4d32a4e9e8dbc82fbbf8d728158f11fbae1b3 |
mtp-v32k.mnn.weight |
80,388,294 | e0e8050043204adf22ec0d5a97d3a8c5c75ee087d35758ddb8f1c1984cbc8da9 |
mtp-v32k.mnn |
33,856 | cd0802dac9e0543d4ca40917cb71e023714fca197ccfeb1319229c8074d1a442 |
License: Apache-2.0, as the original model and conversion.
Credits
- Qwen3.5-2B by the Qwen team (Alibaba Cloud) (Hugging Face). Used under the Apache-2.0 license.
- The MNN conversion (every trunk file here is byte-identical to it) by the taobao-mnn team (Hugging Face). Used under the Apache-2.0 license.
- The MTP head is exported from Qwen/Qwen3.5-2B (Apache-2.0) by TokForge with llmexport from TokForge's fork of MNN.