Gemma 4 31B
Gemma 4 31B is a dense Google model under the Apache 2.0 licence that understands text and images. Quantised, it fits a computer with 32 GB of memory, such as a current Mac or a PC with a graphics card. It does well on German texts in independent tests, which makes it a local all-rounder for office and administration.
Profile
- Maker
- Google · United States
- Licence
- Free for commercial use
- Runs on
- Computer with 32 GB
- Memory at 4 bits
- 21 GB
- Memory at 8 bits
- 33 GB
- Parameters
- 31.3bn
- Context
- 262,000 tokens
- Inputs
- text, image
- Released
- April 2026
Scores
- Human rating
- 1443 · rank 17 of 226 open
- EuroEval
- German: rank 6 of 140
- Epoch index
- 142.9 · rank 29 of 127 open
- GPQA Diamond (expert questions)
- 67.7 %
- AIME (maths)
- 73.3 %
- SimpleQA (factual knowledge)
- 10.4 %
- LMArena · Coding
- 1459 · 25/221
- LMArena · Maths
- 1464 · 15/215
- LMArena · Creative writing
- 1415 · 15/224
- LMArena · Following instructions
- 1433 · 16/226
- LMArena · Hard prompts
- 1448 · 22/226
- LMArena · Long prompts
- 1446 · 14/205
- LMArena · Multi-turn dialogue
- 1452 · 13/224
- LMArena · Building websites (WebDev)
- 1368 · 36/58
- LMArena · Understanding images (Vision)
- 1277 · 3/49
Scores: LMArena, Epoch AI, EuroEval, Hugging Face, models.dev; as of 19 October 2026. Sources and method →
Use cases
Where the model makes sense.
Summarising files and legal documents
- Situation
- A law firm or tax office wants to grasp long letters, notices and contracts faster but may not put confidential documents into a cloud.
- What it looks like
- The model runs on a computer in the office. Staff upload a document and get a summary with the key deadlines, amounts and claims.
- What to watch
- Always check deadlines and figures in the original. The summary does not replace reading; it shows where to look closer.
Reading receipts and delivery notes
- Situation
- Accounting receives invoices, receipts and delivery notes as photos or scans and types them in by hand.
- What it looks like
- The model reads the image and returns supplier, date, amounts and line items as a table or JSON that accounting software can import.
- What to watch
- Errors rise with poor image quality and handwriting. A plausibility check (totals, tax rate) belongs in the process.
Internal knowledge search across manuals
- Situation
- Knowledge sits in manuals, work instructions and minutes, and new staff keep asking the same questions.
- What it looks like
- With an interface such as Open WebUI and document search (RAG), the model answers questions and cites the passage in the document.
- What to watch
- Answers are only as good as the documents. Remove outdated versions first and respect access rights.
Reply drafts in customer service
- Situation
- A service team answers many similar emails and contact-form requests.
- What it looks like
- Based on the request and a few text modules, the model writes a draft in the company’s tone that a person checks and sends.
- What to watch
- Do not send automatically. Customer data stays in-house, and so does responsibility for the answer.
Alt text for images
- Situation
- Hundreds of images on the website and shop lack alt text for accessibility.
- What it looks like
- The model describes each image in one or two sentences; the editorial team accepts or corrects the suggestion.
- What to watch
- Check product and person names; the model does not know your article names.
The use cases were written with AI assistance (Claude by Anthropic) and reviewed by vona. Scores come unchanged from the sources listed.
When another model fits better
For coding and multi-step agents, Qwen 3.8 27B in the same hardware class is a better fit. If answers should come faster on weaker hardware, Gemma 4 26B-A4B is the alternative with similar memory needs.
How to start
ollama run gemma4:31bWith Ollama a single command is enough. The 4-bit version is the usual compromise; if you have enough memory, take the 8-bit version for slightly more accurate results.
Suitable interfaces and programs from our directory: Ollama, LM Studio, Open WebUI
Licence: Free for commercial use
The licence allows commercial use, modification and redistribution; the main obligation is to keep the licence notice.
Hugging Face ↗Similar models
Comparable models.
Google · United States
Gemma 4 26B-A4B
Mixture-of-experts variant of Gemma 4 with four billion active parameters. It responds faster than Gemma 4 31B with similar memory needs.
Free for commercial use · Runs on: Computer with 32 GB
In the overview →Alibaba (Qwen) · China
Qwen 3.8 27B
Compact model from the Qwen 3.8 series that understands images and calls tools. Quantised, it runs on a computer with 32 GB and reaches scores otherwise seen from much larger models.
Free for commercial use · Runs on: Computer with 32 GB
Use cases and details →Mistral AI · France
Mistral Small 3.2
Compact Mistral AI model under the Apache 2.0 licence that runs on a computer with 32 GB. It ranks near the top of EuroEval's German comparison.
Free for commercial use · Runs on: Computer with 32 GB
Use cases and details →
FAQ
What hardware do I need for Gemma 4 31B?
In the common 4-bit version, Gemma 4 31B needs about 21 GB of memory including headroom for context and runtime. That fits: Computer with 32 GB. The more accurate 8-bit version needs about 33 GB. Long documents increase the requirement.
May I use Gemma 4 31B commercially?
The licence allows commercial use, modification and redistribution; the main obligation is to keep the licence notice.
How good is Gemma 4 31B in German?
In EuroEval’s German comparison, Gemma 4 31B ranks 6 of 140. There are no human ratings in German on LMArena yet; test it with your own examples.
AI consulting
Your own language model for your company?
Together we find the model that fits your tasks, your hardware and your data protection requirements, test it with real documents and set up operation.