By GraphicCardPrices. Requirements checked against the official Gemma 4 documentation on .
This guide is for people who want a chat, coding or document assistant that runs on their own computer, with no subscription and no prompts leaving the machine, and who need to know which graphics card is enough.
Gemma 4 is Google DeepMind's open model family. It comes in five sizes: E2B, E4B, 12B, 26B A4B and 31B. The 12B model accepts text, images, audio and video, and the model card lists context windows of 256K tokens for the 12B, 26B A4B and 31B models. The 26B A4B is a mixture-of-experts model that uses about 4B parameters per token, but Google notes that all 26 billion still have to be loaded into memory.
License note: the model card lists the Apache 2.0 license, which permits commercial use, and Google's open source blog announces the Gemma 4 release under Apache 2.0. Read the license page before you build a product on it.
What GPU do you need?
The official documentation says: "Table 1. Approximate GPU or TPU memory required to load Gemma 4 models based on parameter count, quantization level and 20% overhead of loading additional things." (source)
| Tier | VRAM | What it runs | Example cards |
|---|---|---|---|
| Minimum | 6.7 GB | Gemma 4 12B at 4-bit (Q4_0). Google lists 6.7 GB to load it, which leaves little room on an 8GB card for a long conversation. source | RTX 4060, RTX 5060, RTX 3060 (12GB version) |
| Comfortable | 14.4 GB | Gemma 4 26B A4B at 4-bit (Q4_0). The 12B model at 8-bit (SFP8) is listed at 13.4 GB, so a 16GB card covers either. source | RTX 5060 Ti (16GB version), RTX 4060 Ti (16GB version), RTX 5070 Ti |
| Best | 17.5 GB | Gemma 4 31B at 4-bit (Q4_0), the largest size in the family. A 24GB card leaves room for context on top of the weights. source | RTX 3090, RTX 4090, RTX 5090 |
Lowest recorded prices by tier
The cheapest New and Used listings we have recorded for each tier, among NVIDIA GeForce RTX cards with at least that much VRAM. Used includes renewed and refurbished cards. Each is a recorded Amazon US price, not a live checkout price; the model page shows every listing and its condition.
| Tier | Lowest New | Lowest Used |
|---|---|---|
| Minimum (6.7GB+) | $419.99 RTX 5050 8GB Recorded 2026-10-24 11:09:56 UTC | $249.97 RTX 2060 SUPER 8GB Recorded 2026-09-30 05:03:30 UTC |
| Comfortable (14.4GB+) | $788.99 RTX 5060 Ti 16GB Recorded 2026-10-16 21:18:45 UTC | $699.97 RTX 4060 Ti 16GB Recorded 2026-09-29 16:18:54 UTC |
| Best (17.5GB+) | $4,490.00 RTX 4090 24GB Recorded 2026-10-27 18:04:39 UTC | $1,699.99 RTX 3090 24GB Recorded 2026-09-29 16:27:49 UTC |
How to get started
- Pick a size from the table and a precision your card can hold. Google lists each size at 16-bit (BF16), 8-bit (SFP8) and 4-bit (Q4_0).
- For a local chat app, look for the GGUF checkpoints in the Gemma 4 Hugging Face collection. Google's documentation says GGUF checkpoints are available for E2B, E4B, 12B, 26B A4B and 31B, built for drop-in use across the local LLM ecosystem.
- To use it from Python, the model card's Getting Started section installs
transformers,torchandaccelerateand loads the weights withAutoModelForMultimodalLM. - Start with the smallest context window you need, check how much memory your tool really uses, and only then raise it.
Full instructions are in the official Gemma 4 repository.
Things to check before you buy
The table is for the weights only
Google's documentation says the estimates "only account for the memory required to load the static model weights" and "don't include the additional VRAM needed for supporting software or the context window." A long prompt or a long chat uses more memory on top of the table figure, so give yourself margin. Google also notes that the numbers may change with your inference tool.
Less than 6.7 GB?
The same table lists Gemma 4 E4B at 4.5 GB and E2B at 2.9 GB at 4-bit, so smaller cards can run the smaller sizes. Google describes them as models for mobile, edge and browser use, so expect less capability than the 12B.
NVIDIA, AMD and Intel
Google's table does not name a GPU brand. The price table above covers NVIDIA GeForce RTX cards only, because that is what our catalog lists with confirmed memory sizes. Check the documentation of your runtime for AMD or Intel support.
Speed and documentation
Google publishes no consumer-GPU timings in the memory documentation and we have not benchmarked Gemma 4. The documentation page we used also states that it is deprecated in favor of gemma.ai, so check there for newer figures.
Gemma 4 questions
How much VRAM does Gemma 4 need?
Google's documentation lists the memory to load each size. Gemma 4 12B takes 26.7 GB at 16-bit, 13.4 GB at 8-bit and 6.7 GB at 4-bit. The 26B A4B takes 57.7, 28.8 and 14.4 GB, and the 31B takes 69.9, 34.9 and 17.5 GB. These cover the weights only, not the context window.
Can I run Gemma 4 on an 8GB graphics card?
The 12B model at 4-bit is listed at 6.7 GB to load, so it fits an 8GB card with about 1.3 GB left for the context window and your software. The E4B model at 4-bit is listed at 4.5 GB and leaves more room.
Is Gemma 4 free for commercial use?
The model card lists the Apache 2.0 license, which allows commercial use. Read the license page linked from the card before you ship a product.
Compare cards
Estimate memory for language models with the LLM VRAM estimator, see typical prices per memory size in GPU prices by VRAM, or compare pairs: RTX 5060 Ti vs RTX 4060 Ti, RTX 3090 vs RTX 4090, RTX 3060 vs RTX 4060. More local AI GPU guides.
Sources
- Gemma 4 memory requirements table (Google AI documentation), checked
- Gemma 4 12B model card (sizes, context, license, usage), checked
- Gemma 4 license page (linked from the model card), checked
- Gemma 4 and Apache 2.0 (Google Open Source Blog), checked
We have not benchmarked these tools. Requirements change between releases, so check the repository before you buy.