MicrosoftPhi 4Phi3ForCausalLM

Phi 4 Reasoning — Hardware Requirements & GPU Compatibility

ChatMathCodeReasoning

Phi-4-reasoning is Microsoft's reasoning-focused fine-tune of Phi-4, a 14-billion-parameter dense decoder-only transformer, produced through supervised fine-tuning on chain-of-thought traces followed by reinforcement learning, concentrated on math, science, and coding tasks plus safety alignment data. The fine-tuning set combined synthetic prompts with high-quality filtered public web data, and responses are structured in two parts: a reasoning trace followed by a final summarized answer. Microsoft designed it for memory- and compute-constrained, latency-bound scenarios rather than frontier-scale serving, and at 14B parameters it can run on a single high-end consumer GPU once quantized. Context length is 32,768 tokens. It is released under the MIT license, permitting unrestricted commercial and research use, and was published in April 2025.

14.7B params · dense33K contextMITApr 202525.2K downloads231 likes6.8K quant downloads

Based on Phi 4

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs great on a 24 GB GPU

Runs from 8 GB (IQ2_XXS · 16K)Best on 24 GB (Q4_K_M · ~69 tok/s · 32K)Macs from 16 GB (Q4_K_M · ~15 tok/s)

MIT

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Phi 4 Reasoning

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on Tesla P40 (24 GB) (11 GB headroom) · Q4_K_M

Generation speed: ~24 tok/s; generation speed
Cost per 1M output tokens: $1.43; per 1M output tokens

Deployon Vast.aiBest optionfrom $0.12/hr

Compare GPUs →

Run Phi 4 Reasoning

12.5 GB VRAM at 16K context · fits the RTX 4090 (11.5 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/bartowski/microsoft_Phi-4-reasoning-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and search microsoft/phi-4-reasoning, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get microsoft/phi-4-reasoning@q4_k_m
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf bartowski/microsoft_Phi-4-reasoning-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/bartowski/microsoft_Phi-4-reasoning-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf bartowski/microsoft_Phi-4-reasoning-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve microsoft/Phi-4-reasoning --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~32 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model lmstudio-community/Phi-4-reasoning-MLX-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model lmstudio-community/Phi-4-reasoning-MLX-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$bartowski/microsoft_Phi-4-reasoning-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
Tesla P40 (24 GB)CheapestCheapest11 GB free	~24 tok/s	$1.43	Deploy on Vast.aifrom $0.12/hr (opens in new tab)	24 GB · 11 GB free · ~24 tok/s · $1.43 / 1M tokens
RTX A5000 (24 GB)11 GB free	~52 tok/s	$0.85	Deploy on RunPodfrom $0.16/hr (opens in new tab)	24 GB · 11 GB free · ~52 tok/s · $0.85 / 1M tokens
RTX A4000 (16 GB)3 GB free	~31 tok/s	$1.54	Deploy on RunPodfrom $0.17/hr (opens in new tab)	16 GB · 3 GB free · ~31 tok/s · $1.54 / 1M tokens
RTX 5060 Ti 16GB3 GB free	~31 tok/s	$1.70	Deploy on Vast.aifrom $0.19/hr (opens in new tab)	3 GB free · ~31 tok/s · $1.70 / 1M tokens
RTX 4060 Ti 16GB3 GB free	~20 tok/s	$2.67	Deploy on Vast.aifrom $0.19/hr (opens in new tab)	3 GB free · ~20 tok/s · $2.67 / 1M tokens
RTX 4000 Ada (20 GB)7 GB free	~25 tok/s	$2.26	Deploy on RunPodfrom $0.20/hr (opens in new tab)	20 GB · 7 GB free · ~25 tok/s · $2.26 / 1M tokens
RTX A5000 (24 GB)11 GB free	~52 tok/s	$1.13	Deploy on Vast.aifrom $0.21/hr (opens in new tab)	24 GB · 11 GB free · ~52 tok/s · $1.13 / 1M tokens
RTX 3090 (24 GB)11 GB free	~64 tok/s	$0.96	Deploy on RunPodfrom $0.22/hr (opens in new tab)	24 GB · 11 GB free · ~64 tok/s · $0.96 / 1M tokens

Compare GPUs →

How Much VRAM Does Phi 4 Reasoning Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
4.8 GB	11 GB	—	4.3 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 32K context · ~138 tok/s
5.1 GB	11.4 GB	—	4.5 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 32K context · ~128 tok/s
5.3 GB	11.6 GB	—	4.7 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 32K context · ~124 tok/s
5.7 GB	12 GB	—	5.1 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 32K context · ~116 tok/s
6.4 GB	12.7 GB	—	5.8 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 32K context · ~102 tok/s
6.6 GB	12.9 GB	—	5.2 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 32K context · ~100 tok/s
6.8 GB	13.1 GB	—	6.2 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 32K context · ~97 tok/s
7.0 GB	13.2 GB	—	5.5 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 32K context · ~94 tok/s
7.0 GB	13.2 GB	—	6.5 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 32K context · ~94 tok/s
7.1 GB	13.4 GB	—	6.5 GB	3-bit small — 3-bit small quantization	Fits: 32K context · ~92 tok/s
7.3 GB	13.6 GB	—	6.9 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 32K context · ~90 tok/s
7.9 GB	14.2 GB	—	7.4 GB	3-bit medium — 3-bit medium quantization	Fits: 32K context · ~83 tok/s
8.1 GB	14.3 GB	—	8.4 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 32K context · ~81 tok/s
8.2 GB	14.5 GB	—	7.9 GB	3-bit large — 3-bit large quantization	Fits: 32K context · ~80 tok/s
8.6 GB	14.9 GB	—	7.9 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 32K context · ~76 tok/s
9.0 GB	15.3 GB	—	8.4 GB	4-bit small — 4-bit small quantization	Fits: 32K context · ~73 tok/s
9.0 GB	15.3 GB	—	9.3 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 32K context · ~73 tok/s
9.0 GB	15.3 GB	—	8.4 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 32K context · ~73 tok/s
9.5 GB	15.8 GB	—	9.1 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 32K context · ~69 tok/s
9.7 GB	16 GB	—	9.4 GB	4-bit large — 4-bit large quantization	Fits: 32K context · ~68 tok/s
10.8 GB	17.1 GB	—	10.2 GB	5-bit small — 5-bit small quantization	Fits: 32K context · ~61 tok/s
11.2 GB	17.5 GB	—	10.6 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 32K context · ~59 tok/s
11.3 GB	17.6 GB	—	10.9 GB	5-bit large — 5-bit large quantization	Fits: 32K context · ~58 tok/s
12.8 GB	19.1 GB	—	12 GB	6-bit, very good — 6-bit quantization, very good quality	Fits: 32K context · ~51 tok/s
15.4 GB	21.7 GB	—	15.6 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Fits: 32K context · ~43 tok/s
30.0 GB	36.3 GB	—	29.3 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 33 GB

Selected Q4_K_M

4-bit medium quantization

12.5GBat 16K context

Weights: 8.8 GB
KV cache: 3.4 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 11.5 GB free · ~69 tok/s

≥50 tok/s needs ≥732 GB/s · RTX 5090 ~122, RTX 3090 Ti ~69

Which Hardware Can Run Phi 4 Reasoning?

Q4_K_M · 9.5 GB

Phi 4 Reasoning (Q4_K_M) requires 9.5 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 13+ GB is recommended. Using the full 33K context window can add up to 6.3 GB, bringing total usage to 15.8 GB. 40 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080 Ti. 150 devices with unified memory can run Phi 4 Reasoning, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M2 Pro (16 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060IQ2_XXS · 7.7 GB · ~37 tok/s · 16K7.7~3716K8 GB tier →

12 GB · RTX 3060 12GBQ4_K_M · 12.5 GB · ~25 tok/s · 8Ktight12.5~258K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ4_K_M · 12.5 GB · ~31 tok/s · 32K12.5~3132K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 12.5 GB · ~69 tok/s · 32K12.5~6932K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 12.5 GB · ~9 tok/s · 32K12.5~932KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 12.5 GB · ~40 tok/s · 32K12.5~4032KMacs →

How good is Phi 4 Reasoning?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

MMLU-Pro
TIGER-LabIndependent
74.3%
#32/145 open
Independent

Phi 4 Reasoning benchmark results

Alternatives to Phi 4 Reasoning

Phi 4 family · 7 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Apriel 1.6 15B ThinkerServiceNow-AI	—	13.2 GB	~67	same size
Qwen1.5 14BAlibaba	—	22.2 GB	~63	same size
JiRackUltra 14BCMSManhattan	—	12.4 GB	~69	same size
Nemotron Terminal 14BNVIDIA	—	11.9 GB	~69	same size
Ministral 3 14B Reasoning 2512Mistral AI	—	12 GB	~72	same size
Llm Jp 3.1 13B Instruct4llm-jp	—	11.9 GB	~64	same size

Specifications

Publisher: Microsoft
Family: Phi 4
Parameters: 14.7B
Architecture: Phi3ForCausalLM
Context Length: 32,768 tokens
Vocabulary Size: 100,352
Release Date: 2025-04-09
License: MIT

Resources

Frequently Asked Questions

How much VRAM does Phi 4 Reasoning need?

Phi 4 Reasoning requires 9.5 GB of VRAM at Q4_K_M, or 30.0 GB at BF16. Full 33K context adds up to 6.3 GB (15.8 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 14.7B × 4.8 bits ÷ 8 = 8.8 GB

KV Cache + Overhead ≈ 0.7 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 7 GB (at full 33K context)

VRAM usage by quantization

Q4_K_M

9.5 GB

Q4_K_M + full context

15.8 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Phi 4 Reasoning?

Yes, at Q8_0 (15.4 GB) or lower. Higher quantizations like BF16 (30.0 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Phi 4 Reasoning?

For Phi 4 Reasoning, Q4_K_M (9.5 GB) offers the best balance of quality and VRAM usage. Q4_K_L (9.7 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 4.8 GB.

VRAM requirement by quantization

IQ2_XXS

4.8 GB~53%

IQ3_XS

6.8 GB~73%

Q3_K_L

8.2 GB~86%

Q4_K_M ★

9.5 GB~89%

Q4_K_L

9.7 GB~90%

BF16

30.0 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Phi 4 Reasoning on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Phi 4 Reasoning. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 9.5 GB of usable unified memory (RAM minus macOS overhead).

Can I run Phi 4 Reasoning locally?

Yes — Phi 4 Reasoning can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Phi 4 Reasoning?

At Q4_K_M, Phi 4 Reasoning can reach ~504 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~69 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 9.52 × 0.65 = ~546 tok/s

Estimated speed at Q4_K_M (9.5 GB)

NVIDIA B200

~546 tok/s

NVIDIA GeForce RTX 4090

~69 tok/s

NVIDIA B300

~546 tok/s

AMD Instinct MI350X

~504 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Phi 4 Reasoning?

At Q4_K_M, the download is about 8.80 GB. The full-precision BF16 version is 29.32 GB. The smallest option (IQ2_XXS) is 4.03 GB.

Which GPUs can run Phi 4 Reasoning?

40 consumer GPUs can run Phi 4 Reasoning at Q4_K_M. Top options include AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 6900 XT, AMD Radeon RX 6700 XT. 26 GPUs have plenty of headroom for comfortable inference.

Which devices can run Phi 4 Reasoning?

150 devices with unified memory can run Phi 4 Reasoning at Q4_K_M (9.5 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.