Phi 3 Vision 128k Instruct — Hardware Requirements & GPU Compatibility
ChatCodeVision
Phi 3 Vision 128k Instruct is a 4.1B-parameter open language model from Microsoft in the Phi 3 family. It supports a context window of up to 131,072 tokens. At BF16 it needs about 9.40 GB of VRAM — see which GPUs and Macs can run it below.
Runs great on a 24 GB GPU
Runs from 12 GB (BF16 · 8K)Best on 24 GB (BF16 · ~70 tok/s · 32K)Macs from 16 GB (BF16 · ~15 tok/s)
MIT
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Phi 3 Vision 128k Instruct
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
vllm serve microsoft/Phi-3-vision-128k-instruct --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 3090 (24 GB) (9 GB headroom) · BF16
- Generation speed
- ~65 tok/s
- generation speed
- Cost per 1M output tokens
- $0.73
- per 1M output tokens
How Much VRAM Does Phi 3 Vision 128k Instruct Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 9.4 GB | 21.5 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Fits: 32K context · ~70 tok/s |
Selected BF16
Brain floating point 1615.0GBat 16K context
- Weights
- 8.3 GB
- KV cache
- 6.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 9.0 GB free · ~70 tok/s
≥50 tok/s needs ≥722 GB/s · RTX 5090 ~124, RTX 3090 Ti ~70
Which Hardware Can Run Phi 3 Vision 128k Instruct?
BF16 · 9.4 GBPhi 3 Vision 128k Instruct (BF16) requires 9.4 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 131K context window can add up to 50.7 GB, bringing total usage to 60.1 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 3 Vision 128k Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M2 Pro (16 GB).
16–18 GB
16–18 GB tier →fits · 1.0 GB free · up to 16K ctx
18 GPUs · 2 Mac classes · 2 device families
RTX 5080~66Simulate output on RTX 5080RTX 5070 Ti~62Simulate output on RTX 5070 TiRTX 4080 SUPER~51Simulate output on RTX 4080 SUPERM516 GB~11Simulate output on MacBook Air 13" M5 (16 GB)M3 Pro*18 GB~11Simulate output on MacBook Pro 14" M3 Pro (18 GB)iPad~11Simulate output on iPad Pro M5 13" (16 GB)Orin NX~7Simulate output on NVIDIA Jetson Orin NX 16GB20 GB
fits · 5.0 GB free · up to 16K ctx
1 GPU
RX 7900 XT~51Simulate output on RX 7900 XT32–40 GB
fits · 17.0 GB free · up to 32K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~124Simulate output on RTX 5090M5 Max36 GB~34Simulate output on Mac Studio M5 Max (36 GB)M532 GB~11Simulate output on MacBook Air 13" M5 (32 GB)Orin~14Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~9Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at BF16
Alternatives to Phi 3 Vision 128k Instruct
Phi 3 family · 15 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Spark X2.6 4BXHToken | — | 4.3 GB | ~221 | same size |
| NVIDIA Nemotron Labs Dory1.0 4B BF16NVIDIA | — | 5.7 GB | ~214 | same size |
| Granite Vision 4.1 4BIBM | — | 4 GB | ~228 | same size |
| Translategemma 4B ITGoogle | — | 5 GB | ~231 | same size |
| Qwen3 VL 4B InstructAlibaba | — | 4.5 GB | ~208 | same size |
| MiniCPM V 4OpenBMB | — | 3.1 GB | ~236 | same size |
Specifications
- Publisher
- Microsoft
- Family
- Phi 3
- Parameters
- 4.1B
- Architecture
- Phi3VForCausalLM
- Context Length
- 131,072 tokens
- Vocabulary Size
- 32,064
- Release Date
- 2024-05-19
- License
- MIT
Frequently Asked Questions
How much VRAM does Phi 3 Vision 128k Instruct need?
Phi 3 Vision 128k Instruct requires 9.4 GB of VRAM at BF16. Full 131K context adds up to 50.7 GB (60.1 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 4.1B × 16 bits ÷ 8 = 8.3 GB
KV Cache + Overhead ≈ 1.1 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 ≈ 51.8 GB (at full 131K context)
VRAM usage by quantization
Can I run Phi 3 Vision 128k Instruct on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Phi 3 Vision 128k Instruct. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 9.4 GB of usable unified memory (RAM minus macOS overhead).
Can I run Phi 3 Vision 128k Instruct locally?
Yes — Phi 3 Vision 128k Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Phi 3 Vision 128k Instruct?
At BF16, Phi 3 Vision 128k Instruct can reach ~511 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~70 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 ÷ 9.4 × 0.65 = ~553 tok/s
Estimated speed at BF16 (9.4 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Phi 3 Vision 128k Instruct?
At BF16, the download is about 8.29 GB.
Which GPUs can run Phi 3 Vision 128k Instruct?
40 consumer GPUs can run Phi 3 Vision 128k Instruct at BF16. 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 3 Vision 128k Instruct?
150 devices with unified memory can run Phi 3 Vision 128k Instruct at BF16 (9.4 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.