Question 1

How much VRAM does Phi-3.5-MoE-instruct need?

Accepted Answer

Phi-3.5-MoE-instruct (41.9B) requires approximately 93.59239196777344 GB of VRAM at full precision (FP16), 46.8 GB at 8-bit quantization (Q8), and 23.4 GB at 4-bit quantization (Q4). The actual VRAM usage may be slightly higher due to KV cache and runtime overhead.

Question 2

What GPUs can run Phi-3.5-MoE-instruct?

Accepted Answer

We have 10 compatible GPUs indexed for Phi-3.5-MoE-instruct, including H100 NVL, MI300X, RTX Pro 6000 Blackwell, GH200, H200 SXM and 5 more. Compatibility is determined by whether the GPU's VRAM can fit the model's memory footprint, potentially using quantization to reduce memory usage.

Question 3

What is the cheapest GPU to run Phi-3.5-MoE-instruct?

Accepted Answer

The cheapest GPU to run Phi-3.5-MoE-instruct in the cloud is the H100 NVL at $1.80/hr. Other affordable options include MI300X ($1.99/hr) and RTX Pro 6000 Blackwell ($2.09/hr). Prices reflect the lowest on-demand rate across cloud providers.

Question 4

Can Phi-3.5-MoE-instruct run on consumer hardware?

Accepted Answer

Yes, Phi-3.5-MoE-instruct can run on consumer GPUs like the RTX Pro 6000 Blackwell (96 GB VRAM) at full precision (FP16).

Question 5

What quantization options are available for Phi-3.5-MoE-instruct?

Accepted Answer

Phi-3.5-MoE-instruct supports the following quantization levels: FP16 (full precision): ~93.59239196777344 GB, Q8 (8-bit): ~46.8 GB, Q6 (6-bit): ~35.1 GB, Q4 (4-bit): ~23.4 GB. Lower quantization reduces VRAM usage at the cost of some quality loss. Q8 offers a good balance between quality and memory savings, while Q4 provides maximum compression. As a Mixture of Experts (MoE) model with 16 experts (2 active), Phi-3.5-MoE-instruct has a larger total weight size but only activates a subset of parameters per token, which can affect effective memory usage during inference.

GPU	VRAM	Best Quant	From	Rent
H100 NVL	94 GB	Q8	$1.80/hr	Rent
MI300X	192 GB	FP16	$1.99/hr	Rent
RTX Pro 6000 Blackwell	96 GB	Q8	$2.09/hr	Rent
GH200	96 GB	Q8	$2.29/hr
H200 SXM	141 GB	FP16	$2.35/hr	Rent
B200	192 GB	FP16	$5.50/hr
B100	192 GB	FP16	—
MI325X	256 GB	FP16	—
MI350X	288 GB	FP16	—
Gaudi 3	128 GB	FP16	—

Phi-3.5-MoE-instruct

VRAM Requirements

Compatible GPUs

Phi-3.5-MoE-instruct FAQ

VRAM requirements for Phi-3.5-MoE-instruct at different quantization levels
Quantization	VRAM Required
FP16	93.6 GB
Q8	46.8 GB
Q6	35.1 GB
Q4	23.4 GB