AI Potluck
Model components / Base / pretrained models

Phi-4

Microsoft

Microsoft's 14B dense small language model, announced December 2024 and fully released the following month, trained on 9.8 trillion tokens of heavily curated and synthetic data. It is the clearest expression of the small-models-with-high-quality-data thesis, and the line has been extended with mini, multimodal, reasoning and reasoning-vision variants.

Scores base Phi-4 14B; the reasoning variants are separate SKUs. The curated training data is not released, which the openness axis weighs alongside the distribution terms. Verified 2026-08-13 via the microsoft/phi-4 model card and the Artificial Analysis model page.

Openness

3 high confidence
3.0
weights
open(MIT)
data
closed
code
closed
license
MIT(OSI)

MIT weights, which is genuinely permissive with no use restrictions, but neither the training corpus nor the training pipeline is published - Microsoft describes the synthetic data mix without releasing it. So 3, open weights: the artifact is free to use and the recipe is not reproducible. Reaching 4 or 5 would need the data or the pipeline, not a more permissive license.

Adoption

3 high confidence
3.0

626,702 downloads in the trailing 30 days for the single declared artifact microsoft/phi-4, which bands at level 3 (100K-1M) on the model adoption scale.

Capability

3 high confidence
3.0

14B model with outsized math/reasoning: ~56.1% on graduate-level (GPQA-style) science questions, 80.4% on competition math, strong AMC results; reportedly beats GPT-4 (its teacher) on some STEM. Strong-for-size but a small/older model vs 2026 frontier. Artificial Analysis places Phi-4 at #47 of 74 on its Intelligence Index - a mid-table position on the 2026 comparison set, which is what a 3 encodes.

Verified 2026-08-13