Phi-3.5-MoE TPS calculator

Open weights Microsoft 60.8B parameters April 2024

Each card below is assessed against this model at the context length and minimum quality you choose. Speed is an estimate for a single request, calculated from the card's memory bandwidth and the size of the model once compressed.

Calculated for this model

92 of 818 cards that can run it

Smallest card that fits

FirePro S9170

32 GB · Q3_K_M · 26.1 tok/s

Fastest card

B200

310 tok/s · 180 GB

Which GPUs can run Phi-3.5-MoE?

Set the inputs, read the answer

Context length tokens

A longer conversation needs more memory, which can push this model off smaller cards.

Minimum quality

Hides cards that would only fit the model by compressing it below this point.

Search cards

Card memory

Also show cards that cannot run it

92 cards match

Calculating

					Needs	Quantisation	Fit
310 tok/s 186–495 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	62.6 GB	Q8_0	Comfortable
310 tok/s 186–495 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	62.6 GB	Q8_0	Comfortable
247 tok/s 148–396 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	62.6 GB	Q8_0	Comfortable
247 tok/s 148–396 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	62.6 GB	Q8_0	Comfortable
198 tok/s 119–316 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	62.6 GB	Q8_0	Comfortable
195 tok/s 117–312 · low confidence	DRIVE A100 PROD NVIDIA	32 GB	1,870 GB/s	May 2020	27.2 GB	Q3_K_M	Tight
195 tok/s 117–312 · low confidence	GRID A100A NVIDIA	32 GB	1,870 GB/s	May 2020	27.2 GB	Q3_K_M	Tight
189 tok/s 114–303 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	62.6 GB	Q8_0	Comfortable
189 tok/s 114–303 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	62.6 GB	Q8_0	Comfortable
187 tok/s 112–299 · low confidence	GeForce RTX 5090 NVIDIA	32 GB	1,790 GB/s	Jan 2025	27.2 GB	Q3_K_M	Tight
187 tok/s 112–299 · low confidence	GeForce RTX 5090 D NVIDIA	32 GB	1,790 GB/s	Jan 2025	27.2 GB	Q3_K_M	Tight
181 tok/s 109–290 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	62.6 GB	Q8_0	Comfortable
161 tok/s 96–257 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	62.6 GB	Q8_0	Comfortable
161 tok/s 96–257 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	62.6 GB	Q8_0	Comfortable
161 tok/s 96–257 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	62.6 GB	Q8_0	Comfortable
152 tok/s 91–244 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	62.6 GB	Q8_0	Comfortable
139 tok/s 84–223 · low confidence	A100 PCIe 40 GB NVIDIA	40 GB	1,560 GB/s	Jun 2020	34.3 GB	Q4_K_M	Tight
139 tok/s 84–223 · low confidence	A100 SXM4 40 GB NVIDIA	40 GB	1,560 GB/s	May 2020	34.3 GB	Q4_K_M	Tight
139 tok/s 84–223 · low confidence	A800 PCIe 40 GB NVIDIA	40 GB	1,560 GB/s	Nov 2022	34.3 GB	Q4_K_M	Tight
130 tok/s 78–208 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	62.6 GB	Q8_0	Comfortable
130 tok/s 78–208 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	62.6 GB	Q8_0	Tight
130 tok/s 78–208 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	62.6 GB	Q8_0	Comfortable
130 tok/s 78–208 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	62.6 GB	Q8_0	Comfortable
130 tok/s 78–208 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	62.6 GB	Q8_0	Tight
129 tok/s 78–207 · low confidence	GRID A100B NVIDIA	48 GB	1,870 GB/s	May 2020	41.4 GB	Q5_K_M	Tight

Speeds are estimates for a single request — one conversation at a time — calculated from memory bandwidth, model size and quantisation. Real throughput varies with the inference runtime and its version. Figures published by hardware vendors measure many simultaneous requests and are much higher.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation: Microsoft
Organisation type: Industry
Country: United States of America
Published: 23 April 2024
Authors: Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sébastien Bubeck, Martin Cai, Caio César Teodoro Mendes, Weizhu Chen, Vishrav Chaudhary, Parul Chopra, Allie Del Giorno, Gustavo de Rosa, Matthew Dixon, Ronen Eldan, Dan Iter, Amit Garg, Abhishek Goswami, Su…

What it does

The problem areas the model was built for. A model can carry several of each.

Domain: Language
Task: Language modeling/generation, Translation, Question answering, Code generation, Quantitative reasoning

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters: 60.8B
Training data: tokens

Training compute

The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.

Training compute: 3 × 10²³ FLOP
How it was established: Hardware,Operation counting

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware: NVIDIA H100 SXM5 80GB
Chips used: 512
Wall-clock time: 552 hours (23 days)
Power draw: 708.4 kW

Availability

Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.

Weights: Open — downloadable
Model access: Open weights (unrestricted)
Training code: Unreleased
Hugging Face: microsoft

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Likely above 10²³ FLOP: Yes
Record confidence: Confident

Sources

Where this record came from and when it was last checked.

Reference: Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
Last updated: 28 November 2025

The extremes

The ten fastest GPUs for Phi-3.5-MoE

Ranked by estimated tokens per second, newest card first where speeds tie. Because generation is bound by memory bandwidth, this ordering follows bandwidth rather than any gaming benchmark.

The smallest GPUs that still run Phi-3.5-MoE

The cheapest route in, by memory capacity. A tight fit runs the model but leaves nothing spare for a longer conversation.

What the numbers mean

What it takes to run this model

Minimum card

FirePro S9170

Memory needed

27.2 GB

Fastest

310 tok/s

Phi-3.5-MoE sits at 60.8B parameters, which puts it above consumer hardware and into the range where a card is bought for this purpose rather than repurposed for it. 92 of the cards we track can hold it.

The smallest card that holds it is the FirePro S9170 with 32 GB, running it at Q3_K_M and producing around 26.1 tokens per second.

The quickest result comes from a B200 at around 310 tokens per second — its 8,000 GB/s of bandwidth is what buys that.

About this model

Phi-3.5-MoE was published by Microsoft, in United States of America, in April 2024. The organisation is categorised as industry.

It works in Language, and is recorded as doing language modeling/generation, Translation, Question answering, Code generation, Quantitative reasoning.

The weights are published, so it can be downloaded and run on your own hardware indefinitely, offline, with no account attached. It is published under the microsoft organisation on Hugging Face.

How fast it runs, and why

Across every card that can run it, the middle of the range is about 79.0 tokens per second, and 92 of them clear the ten tokens per second that roughly matches reading speed.

This is a mixture-of-experts model, which routes each token through only part of itself. It therefore generates far faster than its total size suggests — while still needing every parameter resident in memory, so it is quick without being cheap to hold.

Without the attention layout on record, the memory column is an approximation. It is close enough to choose hardware by, and least reliable at long context.

Training and provenance

Producing it required around 3 × 10²³ FLOP of arithmetic, on NVIDIA H100 SXM5 80GB, which is a statement about the training budget rather than about inference.

Step by step

How to choose a GPU for Phi-3.5-MoE

The table above has already assessed every card we hold specifications for against this model. Getting to your answer takes six steps.

01

Start from the memory column

Look at what Phi-3.5-MoE actually needs — around 27.2 GB at Q3_K_M. No amount of processing power compensates for a card that cannot hold it.
02

Decide how long your conversations run

The conversation occupies memory too, and grows as it goes. Set the slider to the length you expect: at long context Phi-3.5-MoE can slip off a card that handles short questions easily.
03

Set a quality floor

Compression is what makes Phi-3.5-MoE fit smaller cards, at some cost in accuracy — Q3_K_M on the smallest card that fits. A minimum quality removes the ones that go too far.
04

Compare tokens per second, not specifications

Sort by speed to see how cards rank for Phi-3.5-MoE. It will not match a gaming ordering — generation is bound by memory bandwidth, which is why the B200 tops it at 310 tok/s.
05

Look at the headroom, not just the fit

Tight means Phi-3.5-MoE loads and works, with no room to raise the context later. Comfortable means you can. The difference matters more than a few tokens per second.
06

Check the card from the other side

Following a card through to its own page shows every other model it can hold, which is the question that follows once Phi-3.5-MoE is settled.

Answers

Phi-3.5-MoE — common questions

What is Phi-3.5-MoE used for?

Phi-3.5-MoE works in Language, and is recorded as handling language modeling/generation, Translation, Question answering, Code generation, Quantitative reasoning. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

Where can I download Phi-3.5-MoE?

Its weights are published under the microsoft organisation on Hugging Face. We do not host model files — this site calculates what hardware is needed to run them.

How much compute was used to train Phi-3.5-MoE?

Around 3 × 10²³ FLOP, on NVIDIA H100 SXM5 80GB. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.

Can I run Phi-3.5-MoE if it does not fit in my GPU?

It can be split between the card and system memory, but Phi-3.5-MoE generates painfully slowly that way — the nearest miss we calculate is short by 9.1 GB. Nothing on this page assumes offloading.

Would two GPUs run Phi-3.5-MoE faster?

Two cards buy memory rather than speed. That matters for Phi-3.5-MoE only if one card cannot hold it — 92 can, so a second adds little.

Why does the quantisation differ between cards for Phi-3.5-MoE?

A larger card holds a more accurate copy. Across the cards that run Phi-3.5-MoE, 5 compression levels are used; the floor control above pins it to one.

How accurate are these Phi-3.5-MoE speed estimates?

These are estimates with real error bars. The fastest result here, 186–495 tok/s on the B200, could reasonably land anywhere in its published range depending on which runtime you use.

What GPU do I need to run Phi-3.5-MoE?

The smallest card in our catalogue that holds Phi-3.5-MoE is the FirePro S9170, with 32 GB of memory. It runs the model at Q3_K_M using about 27.2 GB, and produces roughly 26.1 tokens per second. 92 cards in total can run it.

How fast is Phi-3.5-MoE on a GPU?

It depends on the card. The quickest we calculate is a B200 at about 310 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and 92 of the cards that can run Phi-3.5-MoE clear that.

How much VRAM does Phi-3.5-MoE need?

About 27.2 GB at Q3_K_M compression, which is what the smallest card that runs it uses. Less compression needs more: the figures in the memory column above are recalculated for each card, because each one holds the least-compressed version it can.

Is Phi-3.5-MoE open source?

Its weights are published, so Phi-3.5-MoE can be downloaded and run on your own hardware. Note that open weights is not the same as open source in the full sense — it says nothing about the training data, the training code, or the commercial terms attached.

How many parameters does Phi-3.5-MoE have?

Phi-3.5-MoE has 60.8B parameters. "Phi-3.5-MoE has 16x3.8B parameters with 6.6B active parameters when using 2 experts. The model is a mixture-of-expert decoder-only Transformer model using the tokenizer with vocabulary size of 32,064.". That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

Who created Phi-3.5-MoE?

Phi-3.5-MoE was published by Microsoft, based in United States of America, categorised as industry.

When was Phi-3.5-MoE released?

Phi-3.5-MoE was published in April 2024. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

Source

Original publication

Record last updated 28 November 2025

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.

All AI models