Microsoft's small models trained mainly on curated textbook-quality data.
| Feature | Phi | DeepSeek |
|---|---|---|
| Context window | 128k tokens | 128k tokens |
| License | MIT | MIT for code, DeepSeek License for weights |
| Open weights | Yes | Yes |
| Parameters | 3.8B, 14B | 7B to 67B |
| Provider | Microsoft | DeepSeek |
Shaded rows are identical for both; the differing rows are the ones that drive the recommendation below.
Generated from the differences above by structured rules, not by a model.
model · Models
Phi is a family of small language models trained on a deliberately narrow, high-quality corpus of textbooks and exercises rather than on a web-scale crawl. It solves the assumption that capability requires scale, showing that data selection can substitute for a large part of it. It exists because it reframes the question from how much data to what kind of data, and the answer is useful when the budget is a single GPU.
model · Models
DeepSeek is an open-weight model family whose published training methods attracted attention for reaching competitive quality at lower reported cost. It solves the cost question in training rather than in inference, which is where most optimisation effort goes. It exists as evidence that architecture and training-schedule choices matter enough to change what a given budget can produce, and its weights and methods are published.
