LLMs · Microsoft

Phi-3.5 MoE (42B-A6.6B)

Name: Phi-3.5 MoE (42B-A6.6B)
Author: Microsoft

Parameters: 42B (6.6B active)
Architecture: Mixture of experts
Max context: 128K tokens
License: mit
Released: 2024-08

How to run it

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

huggingface.co/lmstudio-community/Phi-3.5-MoE-instruct-GGUF (opens in a new tab)

ollama run hf.co/lmstudio-community/Phi-3.5-MoE-instruct-GGUF

Original weights · Hugging Face

Full-precision files from Microsoft, for fine-tuning, vLLM, Transformers.

huggingface.co/microsoft/Phi-3.5-MoE-instruct (opens in a new tab)

No sign-in needed.

Official repo · GitHub

Reference code, inference scripts and docs from Microsoft. The weights themselves download from Hugging Face.

github.com/microsoft/PhiCookBook (opens in a new tab)

Microsoft’s official Phi cookbook: setup, inference and fine-tuning guides.

git clone https://github.com/microsoft/PhiCookBook

Links checked 2 days ago against Hugging Face, Ollama and GitHub.

Apps that run Phi-3.5 MoE (42B-A6.6B)

Our pick for each kind of computer, assuming the model fits in its memory. Pick your own GPU or device in the Model Finder for an exact answer.

Mac (Apple silicon)

LM Studio

Runs the MLX build, typically 20–35% faster than GGUF for a model with 6.6B active parameters on Apple silicon.

Also good:

Windows PC, NVIDIA GPU

LM Studio

Point-and-click: search, download and chat in one window.

Also good:

Windows PC, AMD Radeon

Lemonade

AMD’s own app: llama.cpp tuned for Radeon on Vulkan and ROCm.

Also good:

Linux, NVIDIA GPU

LM Studio

Point-and-click: search, download and chat in one window.

Also good:

Linux server, 80 GB+ GPUs

vLLM

The full model fits in GPU memory at 8-bit, so it can batch many requests: best for serving a team or an app.

Also good:

App	Type	Opens it as	Systems	Engine	Start
LM Studio	Desktop app	GGUF or MLX	macOSWindowsLinux	llama.cpp, MLX	Open in LM Studio
Ollama	CLI + API	GGUF from Hugging Face	macOSWindowsLinux	llama.cpp, MLX	`ollama run hf.co/lmstudio-community/Phi-3.5-MoE-instruct-GGUF:Q4_K_M`
llama.cpp	CLI + API	GGUF	macOSWindowsLinux	llama.cpp	`llama-server -hf lmstudio-community/Phi-3.5-MoE-instruct-GGUF:Q4_K_M`
Jan	Desktop app	GGUF or MLX	macOSWindowsLinux	llama.cpp, MLX	Open in Jan
MLX LM	CLI + API	MLX	macOS	MLX	`mlx_lm.chat --model mlx-community/Phi-3.5-MoE-instruct-4bit`
vLLM	Server	Original weights	WSLWindowsLinux	vLLM	`vllm serve microsoft/Phi-3.5-MoE-instruct`
SGLang	Server	Original weights	WSLWindowsLinux	SGLang	`python3 -m sglang.launch_server --model-path microsoft/Phi-3.5-MoE-instruct`
Docker Model Runner	CLI + API	GGUF	macOSWindowsLinux	llama.cpp, vLLM	`docker model run hf.co/lmstudio-community/Phi-3.5-MoE-instruct-GGUF:Q4_K_M`
Lemonade	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, ONNX Runtime GenAI	`lemonade pull lmstudio-community/Phi-3.5-MoE-instruct-GGUF:Q4_K_M`
Unsloth Desktop	Desktop app	GGUF or MLX	macOSWindowsLinux	llama.cpp, MLX, stable-diffusion.cpp	Open in Unsloth Desktop
Atomic Chat	Desktop app	GGUF or MLX	macOSWindowsLinux	llama.cpp, MLX-VLM	Open in Atomic Chat
GPT4All	Desktop app	GGUF	macOSWindowsLinux	llama.cpp	Get GPT4All
KoboldCpp	Desktop app	GGUF	macOSWindowsLinux	llama.cpp	Get KoboldCpp
Text Generation WebUI	Desktop app	GGUF	macOSWindowsLinux	llama.cpp, Transformers, ExLlamaV3	Get Text Generation WebUI
llamafile	CLI + API	GGUF	macOSWindowsLinux	llama.cpp	Get llamafile
Msty Studio	Desktop app	GGUF or MLX	macOSWindowsLinux	Ollama, llama.cpp, MLX	Open in Msty Studio
Llama (macOS)	Desktop app	GGUF	macOS	llama.cpp	Open in Llama (macOS)
RecurseChat	Desktop app	GGUF	macOS	llama.cpp	Open in RecurseChat
Inferencer	Desktop app	MLX	macOS	MLX	Get Inferencer
node-llama-cpp	Library	GGUF	macOSWindowsLinux	llama.cpp	Get node-llama-cpp
exo	Server	MLX	macOSLinux	MLX	Get exo
PocketPal AI	Mobile app	GGUF	iOSAndroid	llama.cpp	App Store Google Play
ChatterUI	Mobile app	GGUF	Android	llama.cpp	Download the APK
LLM Farm	Mobile app	GGUF	iOS	llama.cpp	Not on the App Store or TestFlight for now (checked 2026-11-09).

Which apps can open a model depends on the files published for it: GGUF for llama.cpp-based apps, MLX for Apple’s engine, the original weights for vLLM and SGLang. Compare every app.

Memory needed

Weights plus KV cache and runtime overhead. Leave a little extra free for the operating system and other apps. Try another context length.

Quantization	Quality	Weights	Total at 8K context	Total at 32K context	Download
`Q2_K`	Low	16.7 GB	18.4 GB	21.6 GB	GGUF files (opens in a new tab)
`Q3_K_M`	Low	20.5 GB	22.2 GB	25.4 GB	GGUF (Q3_K_L) (opens in a new tab)
`Q4_K_M`	Good	25.5 GB	27.2 GB	30.4 GB	GGUF (opens in a new tab)
`Q5_K_M`	High	29.7 GB	31.4 GB	34.6 GB	GGUF files (opens in a new tab)
`Q6_K`	High	34.3 GB	36.0 GB	39.2 GB	GGUF (opens in a new tab)
`Q8_0`	Very high	44.4 GB	46.1 GB	49.3 GB	GGUF (opens in a new tab)
`F16`	Full	83.7 GB	85.4 GB	88.6 GB	Original weights (opens in a new tab)

Best hardware for Phi-3.5 MoE (42B-A6.6B)

The least expensive GPU in our catalog that runs it without offloading is the AMD Radeon RX 7900 XT (launch price $899) at Q2_K.

Quality	Hardware	Quant	Memory	Speed	Runs
55Good quality (B) · 55 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 79/100 (quality, fit, speed and memory use) sets the ranking.	Apple M4 Max (64 GB)	`Q6_K`	39.8 GB / 48.0 GB	49.8 tok/s	Runs great
55Good quality (B) · 55 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 79/100 (quality, fit, speed and memory use) sets the ranking.	Apple M1 Max (64 GB)	`Q6_K`	39.8 GB / 48.0 GB	38.3 tok/s	Runs great
53Fair quality (C) · 53 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 78/100 (quality, fit, speed and memory use) sets the ranking.	Apple M4 Pro (48 GB)	`Q4_K_M`	30.1 GB / 36.0 GB	38.1 tok/s	Runs great
53Fair quality (C) · 53 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 77/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5090	`Q4_K_M`	28.8 GB / 32.0 GB	157 tok/s	Runs welltight fit
56Good quality (B) · 56 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 77/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA RTX PRO 6000 Blackwell	`Q8_0`	50.9 GB / 96.0 GB	103 tok/s	Runs great
55Good quality (B) · 55 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	Apple M4 Max (128 GB)	`Q6_K`	43.7 GB / 96.0 GB	49.8 tok/s	Runs great
47Fair quality (C) · 47 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	AMD Radeon RX 7900 XTX	`Q3_K_M`	23.4 GB / 24.0 GB	121 tok/s	Runs welltight fit
47Fair quality (C) · 47 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 3090	`Q3_K_M`	23.4 GB / 24.0 GB	130 tok/s	Runs welltight fit
47Fair quality (C) · 47 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4090	`Q3_K_M`	23.4 GB / 24.0 GB	136 tok/s	Runs welltight fit
56Good quality (B) · 56 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 74/100 (quality, fit, speed and memory use) sets the ranking.	Apple M2 Ultra (192 GB)	`Q8_0`	57.6 GB / 144 GB	52.3 tok/s	Runs great

Check it on your own hardware

Run it on a cluster

Devices that can’t run Phi-3.5 MoE (42B-A6.6B) alone can run it together. These example clusters do:

Try it with your own devices

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.