--- license: apache-2.0 base_model: Qwen/Qwen3-0.6B base_model_relation: finetune pipeline_tag: text-generation library_name: mnn language: - en tags: - tokforge - on-device - mnn - android - mobile - speculative-decoding - qwen3 - draft-model - text-generation inference: false --- # Qwen3 0.6B KL baseline 20k draft (MNN) A research draft: Qwen3 0.6B KL-distilled from Qwen3 8B on 20K teacher samples, as an MNN bundle for speculative decoding experiments. Published by [TokForge](https://tokforge.ai) for on-device inference. | | | |---|---| | **What it is** | Research draft (draft model for speculative decoding) | | **Format** | MNN bundle, 4-bit weights (block 128) | | **Runs on** | Android phones (CPU), MNN runtime; not offered in the TokForge app | | **Phone class** | No catalog class; needs more free RAM than its 336 MB of weights | | **Download** | 340 MB (9 files) | | **Base model** | [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) | | **License** | [Apache 2.0](https://www.apache.org/licenses/LICENSE-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 1. Install TokForge from [Google Play](https://play.google.com/store/apps/details?id=dev.tokforge). 2. TokForge does not offer or download this research draft. For speculative decoding with Qwen3 chat models, open **Settings > Speed & engine > Models & Modules** and download **Acceleration Pack: Qwen3-0.6B 40k** ([TokForge-AccelerationPack-Draft](https://huggingface.co/darkmaniac7/TokForge-AccelerationPack-Draft)). 3. To try this draft yourself, run it with MNN's speculative decoding settings below. Collections: [TokForge Research Drafts](https://huggingface.co/collections/darkmaniac7/tokforge-research-drafts-6af3fde6e8cfcf8046db20b6). ## Files | File | Size | What it is | |---|---:|---| | `config.json` | 210 B | MNN runtime settings | | `config_cpu.json` | 167 B | Draft runtime settings (CPU, 2 threads, greedy) | | `export_args.json` | 1.11 KB | Export settings used for the conversion | | `llm.mnn` | 504 KB | Model graph | | `llm.mnn.json` | 1.01 MB | Model graph as JSON (reference copy) | | `llm.mnn.weight` | 336 MB | Model weights | | `llm_config.json` | 4.66 KB | Model settings and chat template | | `onnx/llm.onnx` | 933 KB | ONNX graph from the export (reference) | | `tokenizer.txt` | 1.61 MB | Tokenizer | ## About the model Small `Qwen3-0.6B` draft model exported for **TokForge + MNN** speculative decoding on Android. It is meant to be paired with larger Qwen3 targets; it is not a chat model on its own. - trained on `20K` teacher samples - KL-distilled from a `Qwen3-8B` teacher - exported as a ready-to-use `MNN` bundle - tuned for the classic `CPU draft + GPU/CPU target` TokForge flow | | | |---|---| | **Architecture** | Qwen3 (dense, full attention) | | **Parameters** | 0.6B (4-bit quantized) | | **Quantization** | W4A16 (4-bit weights, block size 128) | | **Teacher** | [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) | ## Recommended settings | Setting | Value | |---------|-------| | Draft model backend | CPU | | Draft threads | 2 | | Draft predict length | d=3 | | Typical target pairing | Qwen3-8B | Typical TokForge recipe: ```json { "backend_type": "opencl", "thread_num": 4, "precision": "low", "memory": "low", "sampler_type": "greedy", "speculative_type": "draftmodel", "draft_predict_length": 3, "draft_config_path": "/path/to/config_cpu.json" } ``` Known-good draft-side config: ```json { "backend_type": "cpu", "thread_num": 2, "precision": "low", "memory": "low", "sampler_type": "greedy" } ``` ## How it was made KL distillation of `Qwen3-0.6B` toward a `Qwen3-8B` teacher on `20K` teacher samples, then exported with MNN's `llmexport` with 4-bit weights in blocks of 128 (`export_args.json` has the full settings). Training acceptance (`alpha`) at the final logged epoch: `0.7178`. ## Limitations - Research draft: TokForge does not offer it, and we do not publish per-device numbers for it. Results vary by device, target model, and workload. - Intended for speculative decoding with larger `Qwen3` targets. - Mobile results depend heavily on backend routing, prompt length, and device thermals. - This is **not** a drop-in Transformers repo for HF Inference; it is a runtime bundle for **TokForge / MNN**. - This is a specialized runtime artifact, not a general-purpose pretrained release. ## License Apache License 2.0 ([text](https://www.apache.org/licenses/LICENSE-2.0)), the license of Qwen3 0.6B and of the Qwen3 8B teacher. ## Credits - Qwen3-0.6B by the Qwen team (Alibaba Cloud) ([Hugging Face](https://huggingface.co/Qwen/Qwen3-0.6B)). Used under the Apache-2.0 license. - Qwen3-8B by the Qwen team (Alibaba Cloud) ([Hugging Face](https://huggingface.co/Qwen/Qwen3-8B)). Used under the Apache-2.0 license. Used as the distillation teacher. - Converted with llmexport from [MNN](https://github.com/alibaba/MNN) by Alibaba (Apache-2.0).