phi-4 is a state-of-the-art open model built to excel at advanced reasoning, coding, and knowledge tasks. It leverages a blend of synthetic data, filtered web data, academic texts, and supervised fine-tuning for precision, alignment, and safety.
发布日期2024年12月12日
参数规模14.7B
上下文长度128K
许可证MIT
知识截止2024年6月1日
Benchmarks
评测成绩
| 评测基准 | 类别 | 分数 | 来源 |
|---|---|---|---|
| 安全防御榜·Toxicity | security | 98.0 | — |
| 安全防御榜·Malicious Use | security | 97.5 | — |
| 安全防御榜·Socioeconomic | security | 96.9 | — |
| 安全防御榜 | security | 96.8 | — |
| 安全防御榜·Bias | security | 96.6 | — |
| 安全防御榜·Child Safety | security | 96.4 | — |
| 安全防御榜·Info Safety | security | 96.2 | — |
| 安全防御榜·Human Rights | security | 96.1 | — |
| MMLU | general reasoning language math | 84.8 | 来源 |
| HumanEval+ | reasoning | 82.8 | 来源 |
| HumanEval | reasoning code | 82.6 | 来源 |
| MGSM | math reasoning | 80.6 | 来源 |
| MATH | math reasoning | 80.4 | 来源 |
| DROP | reasoning math | 75.5 | 来源 |
| Arena Hard | general reasoning creativity | 75.4 | 来源 |
| MMLU-Pro | language reasoning math general | 70.4 | 来源 |
| 价值对齐榜·Data Protection | security | 69.1 | — |
| 价值对齐榜·Morality | security | 67.2 | — |
| 价值对齐榜·Fairness | security | 65.0 | — |
| IFEval | general | 63.0 | 来源 |
| 价值对齐榜 | security | 62.9 | — |
| 价值对齐榜·Legality | security | 58.0 | — |
| PhiBench | reasoning math general | 56.2 | 来源 |
| GPQA | reasoning general | 56.1 | 来源 |
| 价值对齐榜·Safety | security | 55.2 | — |
| LiveBench | math reasoning general | 47.6 | 来源 |
| SimpleQA | general reasoning | 3.0 | 来源 |
Pricing
API 价格对比
| 服务商 | 输入价 | 输出价 | 上下文 | 吞吐(tok/s) | 延迟(s) | 函数调用 | 代码执行 | 联网搜索 |
|---|---|---|---|---|---|---|---|---|
| Azure | $0.13 | $0.50 | 128K | — | — | ✗ | ✗ | ✗ |
| Azure Cognitive Services | $0.13 | $0.50 | 128K | — | — | ✗ | ✗ | ✗ |
价格单位：美元/百万 token，数据来自社区整理，仅供参考。