MNN-Pocket-0.3B
An on-device small-parameter large language model released by Alibaba's MNN team, employing a 300M-parameter dense architecture and supporting INT4 quantization and long context, designed for deployment on memory-constrained devices.
Timeline (last 90 days)
MNN-Pocket-0.3B is a 300M-parameter dense model.
MNN-Pocket-0.3B has a deployment size of 173MB under INT4 quantization.
MNN-Pocket-0.3B achieves a maximum decoding throughput of 466 tokens/s on an iPad Pro M5 (Metal backend).
MNN-Pocket-0.3B supports a context length of up to 32K tokens.
Both the model weights and the MNN deployment package of MNN-Pocket-0.3B are open-sourced.
MNN-Pocket-0.3B was released by the Alibaba MNN team.
The pretraining of MNN-Pocket-0.3B covers approximately 4.1 trillion high-quality Chinese and English tokens.
The pretraining of MNN-Pocket-0.3B uses a four-stage pipeline.
The post-training of MNN-Pocket-0.3B consists of three stages: instruction fine-tuning, multi-teacher on-policy distillation, and quantization-aware training (QAT).
MNN-Pocket-0.3B adopts dependency-aware vocabulary pruning in its architecture design.
2 more timeline events
All Facts (12)
The target devices of MNN-Pocket-0.3B include smartphones and smart glasses.
50%UnverifiedMNN-Pocket-0.3B has a deployment size of 173MB under INT4 quantization.
50%UnverifiedMNN-Pocket-0.3B achieves a maximum decoding throughput of 466 tokens/s on an iPad Pro M5 (Metal backend).
50%UnverifiedMNN-Pocket-0.3B supports a context length of up to 32K tokens.
50%UnverifiedMNN-Pocket-0.3B is a 300M-parameter dense model.
50%UnverifiedMNN-Pocket-0.3B was released by the Alibaba MNN team.
50%UnverifiedThe pretraining of MNN-Pocket-0.3B covers approximately 4.1 trillion high-quality Chinese and English tokens.
50%UnverifiedThe pretraining of MNN-Pocket-0.3B uses a four-stage pipeline.
50%UnverifiedThe post-training of MNN-Pocket-0.3B consists of three stages: instruction fine-tuning, multi-teacher on-policy distillation, and quantization-aware training (QAT).
50%UnverifiedMNN-Pocket-0.3B adopts dependency-aware vocabulary pruning in its architecture design.
50%UnverifiedThe report states that MNN-Pocket-0.3B reaches the frontier level among models of the same size and surpasses some 0.5B-parameter models on instruction-following benchmarks.
50%UnverifiedBoth the model weights and the MNN deployment package of MNN-Pocket-0.3B are open-sourced.
50%