Phi-3.5 MoE (42B-A6.6B) on NVIDIA RTX PRO 5000 Blackwell 72GB

Can the NVIDIA RTX PRO 5000 Blackwell 72GB run Phi-3.5 MoE (42B-A6.6B)?

Yes, it runs great

Runs great. At Q8_0 it needs about 49.7 GB of 72.0 GB usable, with an estimated speed of 85.7 tok/s.

Best quant: Q8_0
Generation speed: 85.7 tok/s
Prompt processing: 7,091 tok/s
First token (1K prompt): 0.3 s
Largest safe context: 128K

Best app to run it on the NVIDIA RTX PRO 5000 Blackwell 72GB

On Windows

LM Studio

Desktop app · runs on llama.cpp, MLX

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

Open Phi-3.5 MoE (42B-A6.6B) in LM Studio

llama.cpp Usually the first to support new model architectures.
Ollama Runs the Hugging Face GGUF with ollama run hf.co/…

On Linux

vLLM

Inference server · runs on 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.

vllm serve microsoft/Phi-3.5-MoE-instruct

LM Studio Point-and-click: search, download and chat in one window.
llama.cpp Usually the first to support new model architectures.

How we choose: the files published for this model, your operating system and GPU, whether it fits fully in memory, and published speed tests (MLX vs llama.cpp on Apple silicon, vLLM vs llama.cpp vs Ollama on NVIDIA). Details.

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

Recommended for the NVIDIA RTX PRO 5000 Blackwell 72GB: Q8_0. Every other quantization has its own download link in the table below.

Quantized files · GGUF

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

Download the Q8_0 file (opens in a new tab)

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:Q8_0

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 3 days ago against Hugging Face, Ollama and GitHub.

How it’s rated

Quality 56Good quality (B) · 56 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking.: How capable the model itself is at Q8_0, from its editorial quality tier scaled by the quantization or precision. It’s the same on any hardware that runs this variant. Scale: A+ 85+ · A 70+ · B 55+ · C 40+ · D under 40.
Runs Runs great: How well it runs on the NVIDIA RTX PRO 5000 Blackwell 72GB: speed, capped by how well it fits in memory. Runs great: 30+ tokens/s (or near the image, video or decision speed target) · Runs well: 10+ tokens/s · Slow: 3+ tokens/s · Very slow: under 3 · Won’t run: doesn’t fit even with offload. A tight fit or RAM offload caps it at Runs well; heavy offload caps it at Slow.

Match score

To order lists, each result also gets a match score out of 100 from four parts. It favors the most capable model that still runs well.

Model qualityEditorial quality tier of the model, scaled by the quantization or precision used.	25.2 / 45
Memory fitFull marks when everything sits in GPU memory; fewer when part spills to system RAM.	25.0 / 25
SpeedGeneration speed against a 40 tokens/s target for LLMs, or render time against a target for image and video.	15.0 / 15
Memory useRewards using the memory you have, so a bigger, better variant beats a tiny one that leaves most memory idle.	15.0 / 15
Match score	80.2 / 100

The match score orders results; it is not shown as a grade. Quality and Runs are the two ratings in the tables.

Every quantization on the NVIDIA RTX PRO 5000 Blackwell 72GB

Quant	Memory needed	Available	Speed	Runs	Quality	Download
`Q2_K`	22.0 GB	72.0 GB	180 tok/s	Runs great	39Basic quality (D) · 39 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking.	GGUF files (opens in a new tab)
`Q3_K_M`	25.8 GB	72.0 GB	158 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 69/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (Q3_K_L) (opens in a new tab)
`Q4_K_M`	30.8 GB	72.0 GB	136 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 73/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q5_K_M`	35.0 GB	72.0 GB	120 tok/s	Runs great	54Fair quality (C) · 54 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 75/100 (quality, fit, speed and memory use) sets the ranking.	GGUF files (opens in a new tab)
`Q6_K`	39.6 GB	72.0 GB	107 tok/s	Runs great	55Good quality (B) · 55 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 77/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q8_0`	49.7 GB	72.0 GB	85.7 tok/s	Runs great	56Good quality (B) · 56 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`F16`	89.0 GB	72.0 GB	11.4 tok/s	Runs welloffloads to RAM	56Good quality (B) · 56 of 100: how capable the model is at F16, whatever it runs on. Match score 58/100 (quality, fit, speed and memory use) sets the ranking.	Original weights (opens in a new tab)

Memory figures assume 8K of context. Longer conversations or documents grow the KV cache.

Other hardware that runs it

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 · 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 · 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 · 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 · 157 tok/s · Runs well (tight 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 · 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 · 49.8 tok/s · Runs great

Quick answers

How much memory does Phi-3.5 MoE (42B-A6.6B) need on the NVIDIA RTX PRO 5000 Blackwell 72GB?

At Q8_0 it needs about 49.7 GB. The NVIDIA RTX PRO 5000 Blackwell 72GB has 72.0 GB usable.

How fast is Phi-3.5 MoE (42B-A6.6B) on the NVIDIA RTX PRO 5000 Blackwell 72GB?

About 85.7 tok/s at Q8_0, an estimate from the hardware’s memory bandwidth and the size of the model.

Which quantization of Phi-3.5 MoE (42B-A6.6B) should I use on the NVIDIA RTX PRO 5000 Blackwell 72GB?

Q8_0. Of the versions that fit, it is the one that runs best; when two run equally well, the check prefers the higher-rated one.

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.