小语言模型

A compact language model can reduce some resource requirements compared with a larger implementation, but the label small alone does not establish a fixed latency, device compatibility, or privacy guarantee. The original Phi-3 report describes a mini model with 3.8 billion parameters trained on 3.3 trillion tokens. Its comparison results belong to the stated benchmarks and testing setup. Phi research also studies carefully filtered and synthetic training data, showing that data choices matter alongside parameter count. Such results should not become a rule that every small model outperforms every large one. Distillation is another approach: a student learns from a teacher’s outputs or distributions under a specified training objective. It does not necessarily inherit every capability of the teacher. Quantization changes numerical representation, often reducing the bits used to store weights. Pruning instead removes or suppresses selected connections, and actual speed benefits depend on whether the implementation exploits the resulting representation. For 3.8 billion scalar weights packed at four bits each, the raw bit count is 15.2 billion, equivalent to 1.9 billion bytes or 1.9 decimal gigabytes. That estimate excludes scales, metadata, activations, caches and software overhead. During autoregressive attention, a key-value cache stores prior attention states to avoid recomputing them for each later token. Longer context can therefore increase runtime memory even when the checkpoint’s parameter count is unchanged. Local deployment is a property of the whole application: logs, synchronization or other requests can still transmit data. Research and deployment comparisons should identify the complete model, runtime, task and measured behavior.