Run gemma-4-31B-it locally: hardware requirements
gemma-4-31B-it needs about 92.4 GB of GPU memory for one person at 32K context, and about 60.0 GB more for each extra person at 64K. For 80 users working as AI agents, plan on 1023 GB, which takes 6 × B200.
Google, 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: google/gemma-4-31B-it
On Hugging FaceGoogle, United States · 31B parameters · License: apache-2.0
What you need for 16 active users
6 × B200
It holds the model and up to 16 active users at 64K context.
The server
2 TB RAM
48 CPU cores
500 GB NVMe
Rules of thumb
- GPU memory in total
- 1023 GB
- Weights
- 62.5 GB (published files)
- Conversation cache, all active
- 960 GB
- Engine (vLLM)
- 3.0 GB
Ways to need less
- Use 32K context per person4 × B200
- Use 8K context per person2 × RTX PRO 6000 Blackwell
- Use a smaller model: gpt-oss-120b1 × B200
- Use a smaller model: Mixtral-8x7B-Instruct-v0.11 × B200
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. 9 cards too small for this model are left out.
| Card | Memory | Active users, one server | For your 16 active |
|---|---|---|---|
| A100 | 80 GB | 9 | 2 servers, each with 8 × A100 |
| B200 | 180 GB | 16 | 6 × B200 |
| H100 | 80 GB | 9 | 2 servers, each with 8 × H100 |
| H200 NVL | 141 GB | 16 | 8 × H200 NVL |
| L40S | 48 GB | 4 | 4 servers, each with 7 × L40S |
| RTX 6000 Ada | 48 GB | 2 | 8 servers, each with 4 × RTX 6000 Ada |
| RTX PRO 5000 Blackwell | 48 GB | 2 | 8 servers, each with 4 × RTX PRO 5000 Blackwell |
| RTX PRO 6000 Blackwell | 96 GB | 8 | 2 servers, each with 6 × RTX PRO 6000 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.
gemma-4-31B-it by team size
AI agents, 1 user in 5 active at a time, 64K context, as published.
| Total users | At once | GPU memory | What you need |
|---|---|---|---|
| 20 | 4 | 303 GB | 2 × B200 |
| 80 | 16 | 1023 GB | 6 × B200 |
| 120 | 24 | 1565 GB | 2 servers, each with 5 × B200 |
| 500 | 100 | 6312 GB | 5 servers, each with 8 × B200 |
Similar models
- diffusiongemma-26B-A4B-itGoogle · 26B · 292 GB for 16 at once
- gemma-4-26B-A4B-itGoogle · 26B · 292 GB for 16 at once
- gemma-3-4b-itGoogle · 4.3B · at least 11.1 GB for 16 at once
- medgemma-4b-itGoogle · 4.3B · at least 11.1 GB for 16 at once
- GLM-4.7-FlashZ.ai · 31B · 437 GB for 16 at once
- NVIDIA-Nemotron-3-Nano-30B-A3B-BF16NVIDIA · 32B · 115 GB for 16 at once
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 gemma-4-31B-it need?
gemma-4-31B-it needs about 92.4 GB of GPU memory for one person at 32K context, and about 60.0 GB more for each extra person at 64K. For 80 users working as AI agents, plan on 1023 GB, which takes 6 × B200.
How many active users can gemma-4-31B-it handle on one 96 GB GPU?
gemma-4-31B-it does not fit on one 96 GB card such as the RTX PRO 6000 Blackwell at 64K context. Compare larger cards or a smaller precision above.
Can gemma-4-31B-it run on a 24 GB GPU such as the RTX 4090?
Not as published for one person at 32K context: it needs about 92.4 GB. A 4-bit or 8-bit build may fit, if the model has one.
What license does gemma-4-31B-it use?
Hugging Face lists its license as apache-2.0. 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