Keimodel
Log inCompare
Azure

Microsoft

Small language models with outsized capability

Microsoft Research's Phi model family proves that size isn't everything. The Phi-4 generation achieves frontier-class performance on reasoning and coding benchmarks with a fraction of the parameters of competing models, enabling fast, cost-effective deployment. Phi models are fully open-weight under the MIT license.

Founded 1975·3 models·Visit Microsoft

3 Models

Phi-4

14B model that punches well above its weight

2 cr
Phi-4 is Microsoft's 14-billion parameter model that achieves remarkable benchmark scores, especially on STEM and reasoning tasks, by training on carefully curated synthetic data rather than raw web text. It outperforms models 5× its size on mathematics and coding.
16k ctxTool useJSON mode

Phi-4 Mini

Ultra-compact model with strong reasoning

2 cr
Phi-4 Mini is a 3.8-billion parameter model that delivers impressive performance on mathematics and code for its size. It is ideal for on-device inference, edge deployments, and latency-critical applications where a larger model would be impractical.
16k ctxTool useJSON mode

Phi-4 Reasoning

Small model with chain-of-thought reasoning

2 cr
Phi-4 Reasoning is Microsoft's reasoning-focused variant, trained with reinforcement learning to produce step-by-step chain-of-thought traces. Despite its compact 14B size, it competes with much larger reasoning models on mathematics and science benchmarks.
33k ctxTool useJSON mode

Other providers

OpenAIOpenAIAnthropicAnthropicGoogleGoogleMetaMeta