// Phi · Text LLM
phi-4: VRAM & local deployment
14.7B parameters. About 12.0 GB at 4-bit and 34.4 GB at BF16 with one 8K-token request.
released 2024-12-11licence: Commercial OKmemory figures are estimates
01Specs
| Publisher | Microsoft · microsoft |
|---|---|
| Family / series | Phi · phi-4 |
| Type | Text LLM (text-generation) |
| Parameters | 14.7B Source: counted from official safetensors weights |
| Context length | 16K tokens (16,384)max_position_embeddings in config.json |
| Architecture | 40 layers · GQA |
| Released weights | BF16 · 29.3 GBdownload |
| Licence | MIT · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed. |
| Released | 2024-12-11 (Hugging Face repo created) |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Smallest GPU | Mac |
|---|---|---|---|---|
| BF16 / FP16 (original) | 29.3 GB | 34.4 GB | 2 × RTX 5090 (64GB) | Mac 64GB |
| FP8 / INT8 / Q8_0 | 15.6 GB | 19.3 GB | RTX 4090 24GB | Mac 36GB |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 8.9 GB | 12.0 GB | 16GB GPU | Mac 24GB |
KV cache per request: 8K 1.7 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- Transformers — mentioned in the official model card
- llama.cpp — official GGUF weights
- Ollama — in the official Ollama library · phi4
ollama run phi404Official repositories & quantised builds
3 official repo(s) from microsoft.
- phi-4 main
- phi-4-gguf GGUF
- phi-4-onnx ONNX
05Other Phi models
All 18 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| Phi-4-reasoning-vision-15B | 15.1B | 12.3 GB | 2026-01-23 |
| Phi-mini-MoE-instruct | 7.6B / 2.4B | 6.6 GB | 2025-06-23 |
| Phi-tiny-MoE-instruct | 3.8B / 1.1B | 3.6 GB | 2025-06-23 |
| Phi-4-mini-flash-reasoning | 3.85B | 4.4 GB | 2025-06-19 |
| Phi-4-mini-reasoning | 3.84B | 4.1 GB | 2025-04-29 |
| Phi-4-reasoning-plus | 14.7B | 12.0 GB | 2025-04-17 |
| Phi-4-reasoning | 14.7B | 12.0 GB | 2025-04-09 |
| Phi-4-multimodal-instruct | 5.57B | 4.2 GB | 2025-02-24 |
06Similar-size Text LLM models from other families
Within ±35% of 14.7B parameters; same estimate method, so the memory columns are directly comparable.
| Model | Family | Params | Context | 4-bit total | Licence | Released |
|---|---|---|---|---|---|---|
| phi-4 | Phi | 14.7B | 16K | 12.0 GB | Commercial OK | 2024-12-11 |
| Qwen3-14B | Qwen | 14.8B | 40K | 11.7 GB | Commercial OK | 2025-04-27 |
| Qwen2.5-14B-Instruct-1M | Qwen | 14.8B | 1M | 12.0 GB | Commercial OK | 2025-01-23 |
| DeepSeek-R1-Distill-Qwen-14B | DeepSeek | 14.8B | 128K | 12.0 GB | Commercial OK | 2025-01-20 |
| Qwen2.5-Coder-14B-Instruct | Qwen | 14.8B | 32K | 12.0 GB | Commercial OK | 2024-11-06 |
| Qwen2.5-14B-Instruct | Qwen | 14.8B | 32K | 12.0 GB | Commercial OK | 2024-09-16 |
| Ministral-3-14B-Instruct-2512-BF16 | Mistral / Magistral / Devstral | 13.9B | 256K | 11.1 GB | Commercial OK | 2025-10-31 |
07Sources
Data checked 2026-07-14. Parameters, context and licence are copied from the official repository; memory figures are our estimates.