Microsoft AIGA
Microsoft Phi-4 (open-weight family)
MIT-licensed small models that punch above their weight — and run on your laptop for $0 per token.
Read the full Microsoft Phi-4 (open-weight family) analysisContext
16K
Max output
—
Input /1M
$0.07
Output /1M
$0.14
Live pricing via OpenRouter
Best for
- On-device and edge inference (3.8B mini in ~3-4GB VRAM)
- Privacy-sensitive / regulated workloads with fully local hosting
- High-volume, well-scoped STEM, extraction, and function-calling tasks
Watch out
Base Phi-4 has only a 16K context window, and the vision model is a math/diagram specialist, not a general multimodal model. Most reasoning/vision benchmarks are vendor-claimed.
For creators. Local document AI, on-device assistants, STEM tutoring, screen/UI automation, and offline drafting — no API cost, full data control.
Benchmarks
| mmlu | 84.8 |
| gpqa | 56.1 |
| math | 80.4 |
| humaneval | 82.6 |
| aime 2025 reasoning plus | 81.3 |
Capabilities
- MIT license — fully open weights, commercial use, fine-tune and redistribute
- Runs on consumer hardware: 14B at Q4 needs ~8-10GB VRAM; 3.8B mini ~3-4GB
- STEM-dense: base 14B beats teacher GPT-4o on GPQA and MATH
- Phi-4-mini: 128K context, built-in function calling
- Phi-4-multimodal: text+vision+speech in 5.6B, 128K context
- Phi-4-reasoning-vision-15B (Mar 2026): adaptive think/no-think
- Runs on Ollama / LM Studio / llama.cpp / ONNX / Foundry Local / HF