Check my GPU

LLMs ·

Gemma 3 12B IT

Parameters
12B
Architecture
Dense
Max context
128K tokens
License
gemma
Released
2025-03

How to run it

Quickest · Ollama

One command downloads a ready-to-run version and starts a chat.

ollama run gemma3:12b

Get Ollama at ollama.com/download (opens in a new tab) · model page (opens in a new tab)

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

huggingface.co/ggml-org/gemma-3-12b-it-GGUF (opens in a new tab)

ollama run hf.co/ggml-org/gemma-3-12b-it-GGUF

Original weights · Hugging Face

Full-precision files from Google, for fine-tuning, vLLM, Transformers.

huggingface.co/google/gemma-3-12b-it (opens in a new tab)

Free Hugging Face account needed: request access on the model page, and Google approves it (usually within minutes to a few hours).

Official repo · GitHub

Reference code, inference scripts and docs from Google. The weights themselves download from Hugging Face.

github.com/google-deepmind/gemma (opens in a new tab)

git clone https://github.com/google-deepmind/gemma

Google DeepMind’s official Gemma library.

Links checked 7 days ago against Hugging Face, Ollama and GitHub.

Memory needed

Weights plus KV cache and runtime overhead. Leave a little extra free for the operating system and other apps.

KV cache: sliding-window hybrid: the KV estimate counts the global-attention layers plus a small allowance for the sliding-window layers.

Quantization Quality Weights Total at 8K context Total at 32K context Download
Q2_K Low 4.9 GB 6.2 GB 8.2 GB GGUF files (opens in a new tab)
Q3_K_M Low 6.0 GB 7.2 GB 9.3 GB GGUF files (opens in a new tab)
Q4_K_M Good 7.4 GB 8.7 GB 10.7 GB GGUF (opens in a new tab)
Q5_K_M High 8.7 GB 9.9 GB 11.9 GB GGUF files (opens in a new tab)
Q6_K High 10.0 GB 11.3 GB 13.3 GB GGUF files (opens in a new tab)
Q8_0 Very high 12.9 GB 14.2 GB 16.2 GB GGUF (opens in a new tab)
F16 Full 24.4 GB 25.7 GB 27.7 GB GGUF (opens in a new tab) Original weights (opens in a new tab)

Best hardware for Gemma 3 12B IT

The least expensive GPU in our catalog that runs it without offloading is the Intel Arc A580 (launch price $179) at Q3_K_M.

Grade · points Hardware Quant Memory Speed Fit
81 ptsGreat (A) · 81 of 100 points: Model quality 27.0/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 14.8/15 AMD Radeon RX 7900 XTX Q8_0 15.4 GB / 24.0 GB 42.3 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 27.0/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 14.8/15 NVIDIA GeForce RTX 3090 Q8_0 15.4 GB / 24.0 GB 47.9 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 27.0/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 14.8/15 NVIDIA GeForce RTX 4090 Q8_0 15.4 GB / 24.0 GB 51.6 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 26.7/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 4070 Ti SUPER Q6_K 12.1 GB / 16.0 GB 46.9 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 26.7/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 4080 SUPER Q6_K 12.1 GB / 16.0 GB 51.4 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 26.7/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 5070 Ti Q6_K 12.1 GB / 16.0 GB 62.6 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 26.7/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 5080 Q6_K 12.1 GB / 16.0 GB 67.0 tok/s Runs well
81 ptsGreat (A) · 81 of 100 points: Model quality 26.7/45, Memory fit 25.0/25, Speed 14.7/15, Memory use 15.0/15 AMD Radeon RX 9070 XT Q6_K 12.1 GB / 16.0 GB 38.5 tok/s Runs well
80 ptsGreat (A) · 80 of 100 points: Model quality 26.2/45, Memory fit 25.0/25, Speed 14.5/15, Memory use 15.0/15 NVIDIA GeForce RTX 5060 Ti 16GB Q5_K_M 10.7 GB / 16.0 GB 37.2 tok/s Runs well
80 ptsGreat (A) · 80 of 100 points: Model quality 25.7/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 4070 SUPER Q4_K_M 9.3 GB / 12.0 GB 50.7 tok/s Runs well

Check it on your own hardware

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.