Phi Microsoft Released 2025-04 reasoningmathsmall

Phi-4 Reasoning

Phi-4 reasoning and reasoning plus are 14B parameter open-weight reasoning models that rival much larger models on complex reasoning tasks.

Best for strong reasoning in a 14B model
Sizes 14B
Context 32K
License MIT
Min VRAM (default size, Q4) 9 GB
Rec VRAM 12 GB

What is Phi-4 Reasoning?

Phi-4 Reasoning is Microsoft's open-weights language model. Released in 2025-04, it ships in 1 size (14B) and is licensed MIT. The most popular use case is strong reasoning in a 14B model.

VRAM and hardware

The 14B size fits in 9 GB of VRAM.

Size Min VRAM (Q4_K_M) Recommended VRAM Tokens/sec on 3090
14B 9 GB 14 GB ~22 tok/s

How to run Phi-4 Reasoning locally

Option 1: Ollama (simplest)

ollama pull phi4-reasoning
ollama run phi4-reasoning

Option 2: Mullama (production)

mullama pull phi4-reasoning
mullama run phi4-reasoning

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "phi4-reasoning gguf")
./llama-cli -m phi4-reasoning.Q4_K_M.gguf -p "Hello, AI!"

Option 4: Python with Mullama or llama-cpp-python

from mullama import Model, Context
model = Model.load("phi4-reasoning.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

What you can build with Phi-4 Reasoning

  • reasoning
  • math
  • small chat

Hardware it fits on

  • Apple Silicon (8GB): 14B in MLX
  • Apple Silicon (16GB): 14B in MLX
  • Apple Silicon (32GB): 14B in MLX or GGUF
  • Apple Silicon (64GB+): 14B in MLX
  • RTX 3090 (24GB): 14B at Q4_K_M
  • RTX 4090 (24GB): same as 3090, ~30% faster
  • RTX 5090 (32GB): 14B at Q8_0
  • 2× RTX 6000 Ada (96GB total): 14B at Q4_K_M

Related models in the Phi family

  • Phi-3 — Mini 3.8B, Medium 14B · small chat, edge deployment
  • Phi-4 — 14B · strong 14B on reasoning benchmarks
  • Phi-4 Mini — 3.8B · small chat with function calling and tool use

Other models you might consider

  • Llama 3.1 (Meta) — 8B, general-purpose chat, agentic workflows, code
  • Llama 3.2 (Meta) — 1B, low-end hardware, mobile, edge deployment
  • Llama 3.2 Vision (Meta) — 11B, vision + chat on desktop GPUs
  • Llama 3.3 (Meta) — 70B, best-in-class 70B for desktops
  • Llama 4 (Meta) — Scout 17B (109B MoE), vision + long context (10M tokens)
  • Qwen 2.5 (Alibaba) — 0.5B, strongest 7B-32B on consumer hardware