Microsoft15B~10 GB VRAM at Q4_K_M

Phi-4 Reasoning Vision 15B — VRAM, speed & local setup

Written by Jakub Rusinowski · Last updated

Microsoft's open multimodal reasoning model: the Phi-4-Reasoning language backbone joined to a SigLIP-2 vision encoder, 15B parameters in all. It takes text and images and decides per request whether to reason step by step before answering. MIT licence; the card states a 16,384-token context.

Phi-4 Reasoning Vision 15B needs about 10 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.

VRAM and speed by quantization

Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.

QuantVRAMSpeed (est.)Fit
Q2_K
2.63 bpw
7.4 GB~101 tok/sFits
Q3_K_M
3.41 bpw
8.9 GB~84 tok/sFits
Q4_K_M
4.83 bpw
11.6 GB~65 tok/sFits
Q5_K_M
5.67 bpw
13.2 GB~57 tok/sFits
Q6_K
6.56 bpw
14.9 GB~51 tok/sFits
Q8_0
8.50 bpw
18.5 GB~41 tok/sFits
F16
16.00 bpw
32.7 GB~3 tok/sOffload

Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Phi-4 Reasoning Vision 15B VRAM calculator →

Get Phi-4 Reasoning Vision 15B running

The cheapest catalogued GPU that runs Phi-4 Reasoning Vision 15B is the Intel Arc B580 (12 GB).

Buy this hardware Intel Arc B580 12GB — 12 GB VRAM · 190 W board powerDeploy in the cloud now RTX 4090 on RunPod — from $0.34/hr · rate checked 2026-07

or compare on Vast.ai from $0.35/hr (typical low · varies)

As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.

Affiliate disclosure: Some links on this page are affiliate links — if you buy through them, LLM Configurator may earn a commission at no extra cost to you. As an Amazon Associate, LLM Configurator earns from qualifying purchases.
Intel Arc B580 12GB
12 GB VRAM · 190 W board power
2026 prices are volatile — check the current listing.

How to run Phi-4 Reasoning Vision 15B

Install Ollama, then run:

ollama run phi-4
Weights on Hugging Face: microsoft/Phi-4-reasoning-vision-15B ↗

Specifications

Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source. Still unconfirmed: context.

Parameters
15 Billion
Context window
16,384
Architecture
Dense transformer (Phi-4-Reasoning) + SigLIP-2 vision encoder, mid-fusion
Provider
Microsoft
Licence
MIT
Availability
Weights downloadable
Practical deployment
Laptop
Specified at
Q4_K_M
System RAM
32 GB
Record updated
2026-11-11
LicenceMITCommercial use permitted

Commercial use permitted. No usage restrictions beyond attribution.

Limits and caveats

  • The card states a 16,384-token context; the config allows 32,768 positions. This page uses the card's figure.

See also Phi-4 (14B)

Quality and use cases

Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.

Best forvisionreasoningmathdocuments

Can I run Phi-4 Reasoning Vision 15B on my GPU?

Other Phi-4 Family sizes

Phi-4 Reasoning Vision 15B — frequently asked questions

How much VRAM does Phi-4 Reasoning Vision 15B need?

About 10 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.

Does Phi-4 Reasoning Vision 15B run on an RTX 4090 (24 GB)?

Yes. Phi-4 Reasoning Vision 15B needs about 10 GB at Q4_K_M, inside a 24 GB card, at an estimated 65 tokens/sec.

How do I run Phi-4 Reasoning Vision 15B locally?

Install Ollama and run `ollama run phi-4`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.

What other sizes does Phi-4 Family come in?

Phi-4 (14B) (9 GB), Phi-4 Reasoning Vision 15B (10 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.