| --- |
| language: en |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.8-27B |
| tags: |
| - mlx |
| - qwen3_5_mtp |
| - speculative-decoding |
| --- |
| |
| # Qwen3.8-27B-MTP-bf16 |
|
|
| This is the MTP (multi-token-prediction) head of [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B). |
|
|
| The official model contains this head as 15 `mtp.*` tensors. |
| We copied these 15 tensors from revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`. |
| We removed the `mtp.` prefix from each tensor name. Example: `mtp.fc.weight` is now `fc.weight`. |
| We did not train the weights. We did not change the weights. |
| The data type is bfloat16, the same as the official release. |
|
|
| ## Contents |
|
|
| - One fusion projection (`fc.weight`). |
| - One transformer layer with its norms. |
| - `config.json` with `model_type: qwen3_5_mtp` and `block_size: 3`. |
| - `model.safetensors.index.json`. Loaders read the head as its own model tree, so these two files are necessary. |
|
|
| ## This is not a standalone model |
|
|
| The head predicts draft tokens from the hidden states of its base model. |
| It cannot generate text alone. |
| Use it with the 4-bit reference model: |
| [EigenLabs/Qwen3.8-27B-4bit](https://huggingface.co/EigenLabs/Qwen3.8-27B-4bit). |
|
|
| ## Role |
|
|
| The Qwen 3.8 MTP challenge uses this head for speculative decoding. |
| The track is `qwen3.8-27b-mtp-v1`. |
| The challenge repository is [Layr-Labs/qwen-3.8-mtp-challenge](https://github.com/Layr-Labs/qwen-3.8-mtp-challenge). |
| The challenge points to one fixed revision of this repository. |
|
|
