GPU VRAM calculator

Run Phi-3-mini-4k-instruct locally: hardware requirements

Phi-3-mini-4k-instruct needs about 11.7 GB of GPU memory for one person at 32K context, and about 24.0 GB more for each extra person at 64K. For 80 users working as AI agents, plan on 34.2 GB, which fits one L40S.

Microsoft, United States. From its files on Hugging Face, checked 2026-11-05.

Context per person: more context lets the model work with longer conversations and more files, and needs more GPU memory for every person. As a guide: 8K for short chat, 32K for chat over documents, 64K for AI agents, whose instructions, tools and files all count.

80 users means about 16 people working at the same moment. That share is a planning assumption; if you know yours, choose the total that matches it.

Work it out for your team: microsoft/Phi-3-mini-4k-instruct

On Hugging Face

Microsoft, United States · 3.8B parameters · License: mit

What you need for 16 active users

1 × L40S

It holds the model and up to 23 active users at 64K context.

The server

128 GB RAM
48 CPU cores
250 GB NVMe

Rules of thumb

GPU memory in total
34.2 GB
Weights
7.6 GB (published files)
Conversation cache, all active
24.0 GB
Engine (vLLM)
3.0 GB

Your options by card

80 users means about 16 using the model at any moment. The others are not waiting; they are simply doing something else. If more become active than a build handles, they queue.

CardMemoryActive users, one serverFor your 16 active
L40S48 GB231 × L40S
RTX 6000 Ada48 GB231 × RTX 6000 Ada
RTX PRO 5000 Blackwell48 GB231 × RTX PRO 5000 Blackwell
A10080 GB431 × A100
H10080 GB431 × H100
RTX PRO 6000 Blackwell96 GB541 × RTX PRO 6000 Blackwell
H200 NVL141 GB821 × H200 NVL
B200180 GB1071 × B200
GeForce RTX 309024 GB232 × GeForce RTX 3090
GeForce RTX 409024 GB232 × GeForce RTX 4090
GeForce RTX 507012 GB82 servers, each with 2 × GeForce RTX 5070
GeForce RTX 508016 GB132 servers, each with 2 × GeForce RTX 5080
GeForce RTX 509032 GB332 × GeForce RTX 5090
RTX 4000 Ada20 GB182 × RTX 4000 Ada
RTX A500024 GB232 × RTX A5000
RTX PRO 4000 Blackwell24 GB232 × RTX PRO 4000 Blackwell
RTX PRO 4500 Blackwell32 GB332 × RTX PRO 4500 Blackwell

These are memory limits. A card can hold more people than it can answer quickly, so test your own workload before you buy. More than one card in a server needs an engine that splits the model.

Phi-3-mini-4k-instruct by team size

AI agents, 1 user in 5 active at a time, 64K context, as published.

Total usersAt onceGPU memoryWhat you need
20416.2 GB1 × RTX 4000 Ada
801634.2 GB1 × L40S
1202446.2 GB1 × A100
500100161 GB1 × B200

All models

How this is worked out

GPU memory = weights + a conversation cache for each person + the serving engine, from this model's files on Hugging Face. These are memory limits, not speed. The full method and its limits.

Questions

How much GPU memory does Phi-3-mini-4k-instruct need?

Phi-3-mini-4k-instruct needs about 11.7 GB of GPU memory for one person at 32K context, and about 24.0 GB more for each extra person at 64K. For 80 users working as AI agents, plan on 34.2 GB, which fits one L40S.

How many active users can Phi-3-mini-4k-instruct handle on one 96 GB GPU?

By memory, about 54 active users at once at 64K context on a card such as the RTX PRO 6000 Blackwell. Speed can set a lower limit, so load-test before you buy.

Can Phi-3-mini-4k-instruct run on a 24 GB GPU such as the RTX 4090?

Yes as published for one person at 32K context: it needs about 11.7 GB.

What license does Phi-3-mini-4k-instruct use?

Hugging Face lists its license as mit. Check the license on the model's page before you use it.

Running a model on your own servers?

NeueCode is private autonomous intelligence for enterprise: it takes on your organization's tasks on your own GPUs, with nothing leaving your network, a person approving every action, and all of it on record. It runs each model on one card, so read the single-card figures above.

What you need to run NeueCode