Phi Mini MoE Instruct — Hardware Requirements & GPU Compatibility
ChatPhi-mini-MoE-instruct is Microsoft's lightweight Mixture-of-Experts model featuring 7.6 billion total and 2.4 billion active parameters. According to the card, it was compressed and distilled from the base model used by Phi-3.5-MoE and GRIN-MoE via the SlimMoE approach, then post-trained with supervised fine-tuning and direct preference optimization for instruction following and safety. It targets English applications in memory- or compute-constrained environments where latency is a concern. Since only 2.4 billion parameters activate per token, it runs quickly, and its full weights fit on one consumer GPU after quantization. The context length reaches 4,096 tokens, which lags behind current standards. Released under the MIT license for both commercial and research use, it came out in June 2025. This model is part of the SlimMoE series, which also houses the smaller Phi-tiny-MoE with 3.8 billion total and 1.1 billion active parameters.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 4K)Best on 24 GB (Q4_K_M · ~239 tok/s · 4K)Macs from 8 GB (IQ3_M · ~42 tok/s)
MIT
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Phi Mini 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/gabriellarson/Phi-mini-MoE-instruct-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 18.6 GB free · 4K context
Download: Phi-mini-MoE-instruct-Q4_K_M.gguf4.99 GBgabriellarson
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 3060 12GB (6 GB headroom) · Q4_K_M
- Generation speed
- ~117 tok/s
- generation speed
- Cost per 1M output tokens
- $0.15
- per 1M output tokens
How Much VRAM Does Phi Mini MoE Instruct Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 3.8 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 4K context · ~281 tok/s | |
| 3.9 GB | — | 3-bit small — 3-bit small quantization | Fits: 4K context · ~278 tok/s | |
| 4.3 GB | — | 3-bit medium — 3-bit medium quantization | Fits: 4K context · ~264 tok/s | |
| 4.4 GB | — | 4-bit · legacy — 4-bit legacy quantization | Fits: 4K context · ~261 tok/s | |
| 5.2 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 4K context · ~239 tok/s | |
| 6.0 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 4K context · ~218 tok/s | |
| 6.9 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Fits: 4K context · ~200 tok/s | |
| 8.2 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 4K context · ~177 tok/s |
Selected Q4_K_M
4-bit medium quantization5.4GBat 4K context
- Weights
- 4.6 GB
- KV cache
- 0.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 18.6 GB free · ~239 tok/s
≥50 tok/s needs ≥134 GB/s · RTX 5090 ~320, RTX 3090 Ti ~239
Which Hardware Can Run Phi Mini MoE Instruct?
Q4_K_M · 5.2 GBPhi Mini MoE Instruct (Q4_K_M) requires 5.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 7+ GB is recommended. Using the full 4K context window can add up to 0.3 GB, bringing total usage to 5.4 GB. 107 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce GTX 1660 SUPER. 172 devices with unified memory can run Phi Mini MoE Instruct, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB).
12 GB
12 GB tier →fits · 6.6 GB free · up to 4K ctx
16 GPUs · 1 device family
RTX 3080 12GB~225Simulate output on RTX 3080 12GBRTX 3080 Ti~225Simulate output on RTX 3080 TiRTX 5070~185Simulate output on RTX 5070iPhone~28Simulate output on Apple iPhone 17 Pro20 GB
fits · 14.6 GB free · up to 4K ctx
1 GPU
RX 7900 XT~103Simulate output on RX 7900 XT32–40 GB
fits · 26.6 GB free · up to 4K ctx
1 GPU · 3 Mac classes · 9 device families
RTX 5090~320Simulate output on RTX 5090M5 Max*36 GB~88Simulate output on Mac Studio M5 Max (36 GB)M632 GB~51Simulate output on iMac 24" M6 (32 GB)RTX Spark N1X~101Simulate output on NVIDIA RTX Spark N1X Laptop (32 GB)Orin~73Simulate output on NVIDIA Jetson AGX Orin 32GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Phi Mini MoE Instruct
Phi family · 5 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| ALoDLM 8BAmazon | — | 7.6 GB | ~119 | same size |
| Ling 3.0 TinyInclusion AI | — | 5.5 GB | ~364 | same size |
| BitCPM CANN 8BOpenBMB | — | 5.6 GB | ~127 | same size |
| Gemma 4 E4BGoogle | — | 5.6 GB | ~123 | same size |
| Olmo 3 7B InstructAllen AI | — | 8.4 GB | ~114 | same size |
| Gemma 3n E4B ITGoogle | — | 7.7 GB | ~127 | same size |
Specifications
- Publisher
- Microsoft
- Family
- Phi
- Parameters
- 7.6B
- Active parameters
- 2.4B 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 Mini MoE Instruct need?
Phi Mini MoE Instruct requires 5.2 GB of VRAM at Q4_K_M, or 15.9 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 7.6B × 4.8 bits ÷ 8 = 4.6 GB
KV Cache + Overhead ≈ 0.6 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 Mini MoE Instruct?
For Phi Mini MoE Instruct, Q4_K_M (5.2 GB) offers the best balance of quality and VRAM usage. Q5_0 (5.3 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 2.7 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Phi Mini MoE Instruct on a Mac?
Yes — Mac Mini M2 (8 GB) and 158 other Macs can run Phi Mini MoE Instruct, the smallest at IQ3_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 5.2 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M2 Pro (16 GB).
Can I run Phi Mini MoE Instruct locally?
Yes — Phi Mini MoE Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Phi Mini MoE Instruct?
At Q4_K_M, Phi Mini MoE Instruct can reach ~149 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~239 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 → 1.6 GB active ÷ (8000 × 0.65) = 0.307 ms, plus 32 layers × 0.055 ms = 1.760 ms, so 1000 ÷ 2.067 ms = ~484 tok/s
Estimated speed at Q4_K_M (5.2 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Phi Mini MoE Instruct?
At Q4_K_M, the download is about 4.59 GB. The full-precision BF16 version is 15.30 GB. The smallest option (IQ2_XXS) is 2.10 GB.
Which GPUs can run Phi Mini MoE Instruct?
107 consumer GPUs can run Phi Mini MoE Instruct at Q4_K_M. Top options include AMD Radeon RX 5700 XT, AMD Radeon RX 580 8GB, AMD Radeon RX 6600, AMD Radeon RX 5600 XT. 95 GPUs have plenty of headroom for comfortable inference.
Which devices can run Phi Mini MoE Instruct?
172 devices with unified memory can run Phi Mini MoE Instruct at Q4_K_M (5.2 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.