Microsoft
Phi-4-mini
Efficient model for low-latency assistance, extraction, and routine automation
Model facts
- Context window
- 128K tokens
- Maximum output
- 4K tokens
- Cheapest paid input
- $0.075 per 1M tokens
- Cheapest paid output
- $0.3 per 1M tokens
- Open weights
- yes
- Providers
- 4
Capabilities and modalities
tool calling, temperature control, open weights, text.
Local hardware estimate
At 4K context: Q4_K_M 3.3 GiB, Q8_0 5.1 GiB, F16 8.5 GiB working memory.
- Parameters
- 3.8 billion
- Architecture
- phi3
- Model license
- mit
Planning estimate adapted from the MIT-licensed whichllm estimator using metadata from Hugging Face; not a vendor minimum requirement.
Artificial Analysis benchmarks
Compare this model on the full LLM benchmark leaderboard.
| Configuration | AAI | Coding | Math | Output tokens/s |
|---|---|---|---|---|
| default | 5.7 | 3.8 | 6.7 | 46.5 |
Observed price history
- 2026-08-24: $0.075 input / $0.3 output per 1M tokens
API providers
- Nvidia — model id microsoft/phi-4-mini-instruct; free input / free output per 1M tokens; provider documentation
- Azure — model id phi-4-mini; $0.075 input / $0.3 output per 1M tokens; provider documentation
- Azure Cognitive Services — model id phi-4-mini; $0.075 input / $0.3 output per 1M tokens; provider documentation
- NanoGPT — model id phi-4-mini-instruct; $0.17 input / $0.68 output per 1M tokens; provider documentation