Phi Tiny MoE Instruct — Hardware Requirements & GPU Compatibility
Chat
Phi Tiny MoE Instruct is a 3.8B-parameter open language model from Microsoft in the Phi family. It supports a context window of up to 4,096 tokens. At Q4_K_M it needs about 2.82 GB of VRAM — see which GPUs and Macs can run it below.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 4K)Best on 24 GB (Q4_K_M · ~309 tok/s · 4K)Macs from 8 GB (Q4_K_M · ~50 tok/s)
MIT
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Phi Tiny MoE 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
ollama run hf.co/tripathyShaswata/Phi-tiny-MoE-instruct-GGUFQ4_K_M fits fully on the RTX 4090 · 20.9 GB free · 4K context
Download: Phi-tiny-MoE-instruct-Q4_K_M.gguf2.25 GBtripathyShaswata
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 3060 12GB (8 GB headroom) · Q4_K_M
- Generation speed
- ~170 tok/s
- generation speed
- Cost per 1M output tokens
- $0.098
- per 1M output tokens
How Much VRAM Does Phi Tiny MoE Instruct Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 2.2 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 4K context · ~346 tok/s | |
| 2.4 GB | — | 3-bit medium — 3-bit medium quantization | Fits: 4K context · ~331 tok/s | |
| 2.8 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 4K context · ~309 tok/s | |
| 3.2 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 4K context · ~289 tok/s | |
| 3.7 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Fits: 4K context · ~272 tok/s | |
| 4.3 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 4K context · ~249 tok/s | |
| 8.1 GB | — | 16-bit · bfloat — Brain floating point 16 — preferred for training | Fits: 4K context · ~167 tok/s |
Selected Q4_K_M
4-bit medium quantization3.1GBat 4K context
- Weights
- 2.3 GB
- KV cache
- 0.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 20.9 GB free · ~309 tok/s
≥50 tok/s needs ≥82 GB/s · RTX 5090 ~386, RTX 3090 Ti ~309
Which Hardware Can Run Phi Tiny MoE Instruct?
Q4_K_M · 2.8 GBPhi Tiny MoE Instruct (Q4_K_M) requires 2.8 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 4+ GB is recommended. Using the full 4K context window can add up to 0.3 GB, bringing total usage to 3.1 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run Phi Tiny MoE Instruct, including NVIDIA DGX Spark, NVIDIA Jetson AGX Thor Developer Kit.
≤ 11 GB
≤ 11 GB tier →fits · 4.9 GB free · up to 4K ctx
16 GPUs · 1 Mac class · 2 device families
RTX 3080~269Simulate output on RTX 3080RTX 2080 Ti~239Simulate output on RTX 2080 TiRTX 3070 Ti~238Simulate output on RTX 3070 TiM3*8 GB~50Simulate output on iMac 24" M3 (8 GB)Orin Nano~61Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)iPhone~38Simulate output on iPhone 1720 GB
fits · 16.9 GB free · up to 4K ctx
1 GPU
RX 7900 XT~119Simulate output on RX 7900 XT32–40 GB
fits · 28.9 GB free · up to 4K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~386Simulate output on RTX 5090M5 Max36 GB~106Simulate output on Mac Studio M5 Max (36 GB)M532 GB~65Simulate output on MacBook Air 13" M5 (32 GB)Orin~111Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~70Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Phi Tiny MoE Instruct
Phi family · 9 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| NVIDIA Nemotron Labs Dory1.0 4B BF16NVIDIA | — | 5.7 GB | ~214 | same size |
| Granite Vision 4.1 4BIBM | — | 4 GB | ~228 | same size |
| Qwen3 4B Instruct 2507Alibaba | — | 4.2 GB | ~226 | same size |
| Qwen2.5 VL 3B InstructAlibaba | — | 3.2 GB | ~249 | same size |
| CAJAL 4BAgnuxo | — | 4.8 GB | ~248 | same size |
| Gemma 4 E4B IT AssistantGoogle | — | 2.7 GB | ~243 | same size |
Specifications
- Publisher
- Microsoft
- Family
- Phi
- Parameters
- 3.8B
- Active parameters
- 1.3B per token
- Architecture
- PhiMoEForCausalLM
- Context Length
- 4,096 tokens
- Vocabulary Size
- 32,064
- Release Date
- 2025-06-23
- License
- MIT
Frequently Asked Questions
How much VRAM does Phi Tiny MoE Instruct need?
Phi Tiny MoE Instruct requires 2.8 GB of VRAM at Q4_K_M, or 8.1 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 3.8B × 4.8 bits ÷ 8 = 2.3 GB
KV Cache + Overhead ≈ 0.5 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 ≈ 0.8 GB (at full 4K context)
VRAM usage by quantization
What's the best quantization for Phi Tiny MoE Instruct?
For Phi Tiny MoE Instruct, Q4_K_M (2.8 GB) offers the best balance of quality and VRAM usage. Q5_K_M (3.2 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 2.2 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Phi Tiny MoE Instruct on a Mac?
Yes — MacBook Air 13" M3 (8 GB) and 145 other Macs can run Phi Tiny MoE Instruct. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 2.8 GB of usable unified memory (RAM minus macOS overhead).
Can I run Phi Tiny MoE Instruct locally?
Yes — Phi Tiny MoE Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Phi Tiny MoE Instruct?
At Q4_K_M, Phi Tiny MoE Instruct can reach ~152 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~309 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 1.0 GB active ÷ (8000 × 0.65) = 0.186 ms, plus 32 layers × 0.055 ms = 1.760 ms, so 1000 ÷ 1.946 ms = ~514 tok/s
Estimated speed at Q4_K_M (2.8 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Phi Tiny MoE Instruct?
At Q4_K_M, the download is about 2.25 GB. The full-precision BF16 version is 7.51 GB. The smallest option (Q2_K) is 1.60 GB.
Which GPUs can run Phi Tiny MoE Instruct?
52 consumer GPUs can run Phi Tiny MoE Instruct at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.
Which devices can run Phi Tiny MoE Instruct?
163 devices with unified memory can run Phi Tiny MoE Instruct at Q4_K_M (2.8 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.