Check my GPU

LLMs ·

Gemma 4 26B-A4B

Parameters
26B (3.8B active)
Architecture
Mixture of experts
Max context
256K tokens
License
apache-2.0
Released
2026-03

How to run it

Quickest · Ollama

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

ollama run gemma4:26b

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/unsloth/gemma-4-26B-A4B-it-GGUF (opens in a new tab)

ollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF

Original weights · Hugging Face

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

huggingface.co/google/gemma-4-26B-A4B-it (opens in a new tab)

No sign-in needed.

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 5 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 10.3 GB 11.3 GB 12.5 GB GGUF (UD-Q2_K_XL) (opens in a new tab)
Q3_K_M Low 12.6 GB 13.6 GB 14.9 GB GGUF (opens in a new tab)
Q4_K_M Good 15.7 GB 16.7 GB 18.0 GB GGUF (opens in a new tab)
Q5_K_M High 18.3 GB 19.3 GB 20.5 GB GGUF (opens in a new tab)
Q6_K High 21.2 GB 22.2 GB 23.4 GB GGUF (opens in a new tab)
Q8_0 Very high 27.4 GB 28.4 GB 29.6 GB GGUF (opens in a new tab)
F16 Full 51.6 GB 52.6 GB 53.8 GB GGUF · folder (opens in a new tab) Original weights (opens in a new tab)

Best hardware for Gemma 4 26B-A4B

The least expensive GPU in our catalog that runs it without offloading is the Intel Arc B580 (launch price $249) at Q2_K.

Grade · points Hardware Quant Memory Speed Fit
96 ptsExcellent (S) · 96 of 100 points: Model quality 41.6/45, Memory fit 25.0/25, Speed 14.7/15, Memory use 15.0/15 Apple M4 Max (64 GB) Q8_0 32.2 GB / 48.0 GB 38.7 tok/s Runs well
96 ptsExcellent (S) · 96 of 100 points: Model quality 41.2/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 5090 Q6_K 23.8 GB / 32.0 GB 201 tok/s Runs well
95 ptsExcellent (S) · 95 of 100 points: Model quality 41.6/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 13.8/15 NVIDIA RTX PRO 6000 Blackwell F16 57.4 GB / 96.0 GB 67.9 tok/s Runs well
95 ptsExcellent (S) · 95 of 100 points: Model quality 41.2/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15 AMD Radeon RX 7900 XTX Q6_K 23.4 GB / 24.0 GB 92.6 tok/s Tight fit
95 ptsExcellent (S) · 95 of 100 points: Model quality 41.2/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 3090 Q6_K 23.4 GB / 24.0 GB 105 tok/s Tight fit
95 ptsExcellent (S) · 95 of 100 points: Model quality 41.2/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 4090 Q6_K 23.4 GB / 24.0 GB 113 tok/s Tight fit
93 ptsExcellent (S) · 93 of 100 points: Model quality 41.2/45, Memory fit 25.0/25, Speed 14.7/15, Memory use 12.5/15 Apple M1 Max (64 GB) Q6_K 26.0 GB / 48.0 GB 38.6 tok/s Runs well
91 ptsExcellent (S) · 91 of 100 points: Model quality 39.5/45, Memory fit 25.0/25, Speed 14.6/15, Memory use 12.6/15 Apple M4 Pro (48 GB) Q4_K_M 19.6 GB / 36.0 GB 38.0 tok/s Runs well
90 ptsExcellent (S) · 90 of 100 points: Model quality 41.6/45, Memory fit 25.0/25, Speed 14.7/15, Memory use 8.7/15 Apple M4 Max (128 GB) Q8_0 36.0 GB / 96.0 GB 38.7 tok/s Runs well
88 ptsExcellent (S) · 88 of 100 points: Model quality 35.0/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15 AMD Radeon RX 9070 XT Q3_K_M 14.4 GB / 16.0 GB 115 tok/s Tight fit

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.