Main Model •
Collection •
GitHub (MNN) •
Tech Report
English | 中文
MNN-Pocket-0.3B-Base
The pre-trained base checkpoint of MNN-Pocket-0.3B, a ~300M-parameter compact LLM co-designed with the MNN inference stack for on-device deployment. It is the starting point for continued pre-training and fine-tuning; for chat / assistant use take the instruction-tuned MNN-Pocket-0.3B instead.
Find more details in the original model card: https://huggingface.co/taobao-mnn/MNN-Pocket-0.3B
- Trained with a multi-stage, jointly-searched data recipe (domain x quality x difficulty x density x style) under a fixed tokens-per-parameter budget
- Competitive at the ~300M scale: average 33.58 on the 11-benchmark base suite, ahead of LFM2.5-350M-Base, SmolLM2-360M-Base and Gemma-3-270M-Base (full table in the tech report)
- Same architecture as the instruct model, so any standard Hugging Face tooling works out of the box
Demo
iPhone 18 Pro Demo
Your browser does not support the video tag.Model Details
| Property | Value |
|---|---|
| Checkpoint | Pre-trained base (no instruction tuning) |
| Parameters | ~300M |
| Architecture | Decoder-only transformer, 28 layers |
| Hidden size | 896 |
| Attention | 14 heads, GQA with 2 KV heads, head_dim 64 |
| FFN intermediate | 2560 (SiLU) |
| Context Length | 32K (40,960 position embeddings) |
| Vocabulary | 65,536 |
| Embeddings | tied input/output |
| Precision | bfloat16 |
| Languages | Chinese, English |
🏃 How to run
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "taobao-mnn/MNN-Pocket-0.3B-Base"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_id)
# base checkpoint: plain text completion, no chat formatting
For fine-tuning, see the Fine-Tuning section of the main model card.
Other Formats
- MNN-Pocket-0.3B — BF16 instruction-tuned, full model card (723 MB)
- MNN-Pocket-0.3B-SFT — supervised fine-tuning checkpoint
- MNN-Pocket-0.3B-GGUF / MNN-Pocket-0.3B-MNN — GGUF and INT4 MNN deployment builds
- MNN-Pocket-0.3B-Dec — non-autoregressive decision model (LoRA adapter + pointer head)
License
Apache License 2.0. See LICENSE.txt.