Microsoft’s compact model, surprisingly capable at reasoning and code
Phi-4
Phi-4 is a 14-billion parameter language model developed by Microsoft Research. It is part of the Phi model family, emphasizing data quality and reasoning capabilities over pure scale.
Intended uses
Phi-4 is designed for:
- STEM reasoning and problem-solving (math, physics, coding, logic)
- Memory and compute-efficient applications
- Educational and research applications
- Conversational AI and chatbot experiences
Characteristics
| Attribute | Details |
|---|---|
| Provider | Microsoft |
| Architecture | phi3 |
| Cutoff date | June 2024 |
| Languages | English (primary), German, Spanish, French, Portuguese, Italian, Hindi, Japanese |
| Tool calling | ❌ |
| Input modalities | Text |
| Output modalities | Text |
| License | MIT |
Available model variants
| Model variant | Parameters | Quantization | Context window | VRAM¹ | Size |
|---|---|---|---|---|---|
ai/phi4:latestai/phi4:14B-Q4_K_M | 15B | IQ2_XXS/Q4_K_M | 16K tokens | 9.78 GiB | 8.43 GB |
ai/phi4:14B-Q4_0 | 15B | Q4_0 | 16K tokens | 9.16 GiB | 7.80 GB |
ai/phi4:14B-Q4_K_M | 15B | IQ2_XXS/Q4_K_M | 16K tokens | 9.78 GiB | 8.43 GB |
ai/phi4:14B-F16 | 15B | F16 | 16K tokens | 27.97 GiB | 27.31 GB |
¹: VRAM estimated based on model characteristics.
latest→14B-Q4_K_M
Use this AI model with Docker Model Runner
First, pull the model:
bashdocker model pull ai/phi4
Then run the model:
bashdocker model run ai/phi4
For more information on Docker Model Runner, https://docs.docker.com/desktop/features/model-runner/.
***ations
- Phi-4 is optimized for single-turn queries rather than long multi-turn conversations.
- Hallucinations may occur, particularly in factual knowledge recall.
- Instruction-following capabilities are less robust than some larger models.
Benchmark performance
| Category | Benchmark | phi-4 | phi-3 |
|---|---|---|---|
| Popular Aggregated Benchmark | MMLU | 84.8 | 77.9 |
| Science | GPQA | 56.1 | 31.2 |
| Math | MGSM | 80.6 | 53.5 |
| MATH | 80.4 | 44.6 | |
| Code Generation | HumanEval | 82.6 | 67.8 |
| Factual Knowledge | SimpleQA | 3.0 | 7.6 |
| Reasoning | DROP | 75.5 | 68.3 |
Links
- Introducing Phi-4
- https://arxiv.org/abs/2412.08905
ai 命名空间的其他可用镜像:
ai/qwen3
Qwen3 大语言模型,适合 AI 应用开发
ai/deepseek-r1-distill-llama
DeepSeek R1 Distill Llama,大语言模型
ai/qwen3-embedding
Docker 官方模型镜像,Qwen3 文本嵌入与重排序,适合 RAG、语义检索与向量化流水线
ai/qwen3-vl
Docker 官方 Qwen3 多模态(VL)模型镜像,适合图文理解、视觉问答与多模态应用
ai/qwen3-vllm
Docker 官方 Qwen3 与 vLLM 推理组合镜像,适合高吞吐文本生成与 OpenAI 兼容 API 服务
ai/qwen3-embedding-vllm
Docker 官方 Qwen3 Embedding + vLLM 推理镜像,适合高吞吐语义检索与 RAG 向量化服务
镜像拉取方式
您可以使用以下命令拉取该镜像。请将 <标签> 替换为具体的标签版本。如需查看所有可用标签版本,请访问 标签列表页面。
DockerHub 原生拉取命令
docker model pull ai/phi4:<标签>
用户好评
来自真实用户的反馈,见证轩辕镜像的优质服务