Gemma 4 31B

About

Google aimed this size of Gemma 4 at maximum quality and as a base for further training. It can take a whole code repository or a long document in one request. Text and pictures can be mixed in any order within the same request.

How it compares

In LM Arena's public vote for reading pictures, it ranks second of 15 open models.

In LM Arena's public vote for coding, it ranks fifth of 59 open models.

Of the 7 chat models Outmake recommends for a 32 GB Mac, Outmake estimates it is the slowest, with Qwen3.6 35B about six times as fast.

Limitations

It can state facts that are wrong or out of date.

Its training data stops in January 2025.

It can miss sarcasm, subtle nuance or figurative language.

Good to know

It was trained on text in more than 140 languages.

Certain personal information was filtered out of its training data before training.

Sources: google/gemma-4-31B-it model card, Gemma 4: Our most capable open models to date, Outmake's speed estimates, LM Arena and Hugging Face. Every line comes from one of these.

Run It Anywhere

Gemma 4 31B is an open model, so any app that runs GGUF files can run it. These take the Q4_K_M size from unsloth straight from Hugging Face.

• LM StudioMac, Windows, LinuxOpen
• JanMac, Windows, LinuxOpen
• MstyMac, Windows, LinuxOpen
• OllamaMac, Windows, Linuxollama run hf.co/unsloth/gemma-4-31B-it-GGUF:Q4_K_MCopy
• llama.cppMac, Windows, Linuxllama-server -hf unsloth/gemma-4-31B-it-GGUF:Q4_K_MCopy
• Docker Model RunnerMac, Windows, Linuxdocker model run hf.co/unsloth/gemma-4-31B-it-GGUF:Q4_K_MCopy