Mixture of experts: big-model quality at small-model speed. Made by Google.
Gemma 4 26B A4B is a general chat model from Google. Mixture of experts: big-model quality at small-model speed. Running it locally needs about 19.4 GB of memory, and the download is 18 GB. It is a mixture of experts model: only about 2.8 GB of its weights are read for each word, so it generates much faster than a dense model of the same size, but all 18 GB still has to fit in memory.
The least memory you need, by Apple chip:
Runs well on: RTX 3090, RTX 4090, RTX 5090, Radeon RX 7900 XTX. Runs slowly with help from system RAM on: RTX 3060 12GB, RTX 3070, RTX 3080 10GB, RTX 4060, RTX 4060 Ti 16GB, RTX 4070 / 4070 Super, RTX 4070 Ti Super, RTX 4080 / 4080 Super, RTX 5060, RTX 5060 Ti 16GB, RTX 5070, RTX 5070 Ti, RTX 5080, Radeon RX 7800 XT, Radeon RX 9070 XT.
As an Amazon Associate I earn from qualifying purchases.
About 19.4 GB of video memory or unified memory (the download is 18 GB, plus room for the conversation). On a Mac, the smallest configuration that runs it is 24 GB.
A 16 GB M4 Mac cannot run it. An 8 GB M1 Mac cannot run it.
Yes, the model is free to download. Its licence is listed as Open weights. Read the licence terms before using it in a business.
Install Ollama, then run: ollama run gemma4:26b
| Machine | Verdict | Speed |
|---|---|---|
| M1 8GB | Too big | |
| M4 16GB | Too big | |
| M4 32GB | Runs great | 14 words/s |
| M4 Pro 48GB | Runs great | 31 words/s |
| M4 Max 128GB | Runs great | 61 words/s |
| M3 Ultra 512GB | Runs great | 92 words/s |
| RTX 4060 | Slow | 9 words/s |
| RTX 4070/4070 Super | Slow | 13 words/s |
| RTX 5070 Ti | Slow | 23 words/s |
| RTX 3090 | Runs great | 89 words/s |
| RTX 5090 | Runs great | 172 words/s |
| Ryzen AI Max+ 395 128GB | Runs great | 25 words/s |
| CPU only 32GB | Slow | 6 words/s |