Phi-4
Phi-4 is an open-weights language model from Microsoft, released December 12, 2024. It has 15B parameters (dense). The context window is 16K tokens. On the Epoch Capabilities Index it scores 129.0, rank 159 of 256 rated models.
Check it against your own PC →
Can you run Phi-4 locally?
The smallest common GPU it runs on is 8 GB of VRAM (with 32 GB of RAM): IQ4_XS, split GPU / CPU, roughly 6 tokens per second. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.
| GPU memory | Verdict | Best quant | VRAM | RAM | tok/s |
|---|---|---|---|---|---|
| 6 GB VRAM | Very slow | Q3_K_M (okay) | 5.2 GB | 5.9 GB | ~5 |
| 8 GB VRAM | Runs, slowly | IQ4_XS (good) | 7.2 GB | 4.5 GB | ~6 |
| 12 GB VRAM | Runs on GPU | Q3_K_M (okay) | 11.1 GB | – | ~29 |
| 16 GB VRAM | Runs on GPU | Q5_K_M (excellent) | 14.4 GB | – | ~28 |
| 24 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 16.0 GB | – | ~45 |
| 32 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 16.0 GB | – | ~82 |
Ready-made quantizations on Hugging Face: MaziyarPanahi/phi-4-GGUF · bartowski/phi-4-GGUF · lmstudio-community/phi-4-GGUF · unsloth/phi-4-GGUF · mradermacher/phi-4-i1-GGUF · EasierAI/Phi-4-14B
Strengths by category
ModelTape's 0–100 fit on Epoch AI benchmark results.
| Category | Score | Based on |
|---|---|---|
| Math | 17.0 | MATH level 5, OTIS Mock AIME 2024-2025 |
| Knowledge | 39.4 | GPQA diamond, MMLU |
Benchmark results
| Benchmark (Epoch AI) | Score |
|---|---|
| MMLU | 79.7% |
| MATH level 5 | 64.9% |
| GPQA Diamond | 43.1% |
| OTIS Mock AIME | 13.7% |
| Balrog | 11.6% |
Links
Hugging Face: microsoft/phi-4 · Epoch AI benchmarks
Models with a similar score
Mistral Medium 3 · Qwen3-8B · Llama 4 Maverick · DeepSeek-V3 · Qwen2.5-Max · Gemini 1.5 Pro (Sept 2024) · Claude 3.5 Sonnet · Grok-2 (Dec 2024)
More from Microsoft
MAI-Cyber-1.1-Flash · MAI-Base-2 · MAI-Cyber-1-Flash · MAI-Code-1.1-Flash · Mage VL · MAI-Thinking-1 · MAI-Code-1-Flash · phi-3-small 7.4B