Gemma 4 Chat
Where to talk to Gemma 4 for free, how to chat privately on your own computer or phone, and the settings that give the best answers.
This site does not host a chatbot - it points you to the official places to chat ·
Where to chat with Gemma 4
Two choices matter most: how much setup you will accept, and whether your messages may leave your device.
| Where | Setup | Models | Your messages | Cost |
|---|---|---|---|---|
| Google AI Studio | Sign in with a Google account | 31B, 26B A4B | Sent to Google | Free to try |
| Ollama | Install, then one command | All five sizes | Stay on your computer | Free |
| LM Studio | Desktop app, no terminal | All five sizes | Stay on your computer | Free |
| Google AI Edge Gallery | Phone app plus a model download | E2B, E4B | Stay on your phone | Free |
| OpenRouter | Create an account | 31B, 26B A4B | Sent to a hosting provider | Pay per token |
Current per-token prices are on the pricing page. To compare prompts side by side before installing anything, see the playground guide.
Start chatting in minutes
Pick the route that matches your device.
Open Google AI Studio
Go to aistudio.google.com and sign in with a Google account.
Choose a Gemma 4 model
In the model list, pick Gemma 4 31B for the best answers or 26B A4B for faster ones.
Type and send
Add a system instruction if you want a particular tone, attach an image if you have one, and start chatting.
Nothing to download. Your messages are processed on Google's servers, so use a local option for anything confidential.
ollama run gemma4
That downloads the default build (the same as gemma4:e4b) and opens a chat in your terminal. For a 16 GB machine, the 12B is a better conversationalist:
ollama run gemma4:12b-it-qat
ollama run gemma4:12b-it-qat
# Inside the chat:
/set system You are a concise assistant.
/set parameter num_ctx 32768
/clear # forget the conversation so far
/bye # leave
Get Ollama from ollama.com/download. Every tag and download size is in the download hub; step-by-step setup is in the install guide.
Install LM Studio
Download it from lmstudio.ai for Windows, macOS or Linux.
Download a Gemma 4 model
Press Ctrl + Shift + M (Cmd on Mac), search "Gemma" and pick a size that fits your memory.
Load it and chat
Press Ctrl / Cmd + L, select the model and start typing. No terminal needed.
On Apple silicon, LM Studio can use MLX builds - see Mac setup.
Install Google AI Edge Gallery
Get it from Google Play.
Download E2B or E4B
E2B is a 2.58 GB download and E4B is 3.66 GB. Use Wi-Fi.
Chat offline
Once the model is on the phone, chat works with no connection, and nothing you type leaves the device.
Full walkthrough on the Android page; other phones are covered in mobile downloads.
Which model to chat with
Bigger models hold longer, more accurate conversations. Smaller ones answer faster on modest hardware.
| You have | Chat with | Memory at 4-bit | What to expect |
|---|---|---|---|
| Just a browser | 31B in AI Studio | None | The strongest answers, online only |
| A phone | E2B or E4B | 1.1 - 2.5 GB | Quick everyday help, voice and photos, offline |
| A laptop with 8 GB | E4B | 4.5 GB | Good general chat; weaker on long reasoning |
| A laptop with 16 GB | 12B | 6.7 GB | The best balance, with text, image and audio input |
| A 24 GB GPU or 32 GB Mac | 26B A4B or 31B | 14.4 - 17.5 GB | Near the top of the family; the 26B answers faster |
Memory figures are Google's estimates for loading the weights; a long conversation needs more on top. Plan with the VRAM guide, compare sizes on the models page and see the 12B guide. Prefer the QAT builds when you run locally.
Recommended chat settings
Google's model card gives one set of sampling values for every size. Most apps already use them; check if answers seem off.
How adventurous the wording is.
Keeps the most likely 95% of choices.
Considers the 64 likeliest next words.
System instruction
Gemma 4 supports a real system role. Use it once, at the start, for tone, format and rules that should hold for the whole conversation.
Thinking
Turn it on for maths, code and planning; leave it off for quick back-and-forth. It works best set once per conversation, not switched every turn.
Context length
The models take 128K or 256K tokens, but local apps often start with a much smaller window. Raise it if the chat forgets earlier messages.
From Google's model card and prompt formatting guide. How thinking works is explained on the capabilities page.
Getting better answers
One task per message
Ask for one thing at a time and build on the answer, especially with the small models.
Say what you want back
Name the format, length and audience: "three bullet points for a beginner" beats "explain this".
Paste in current facts
Its training data ends in January 2025. For anything newer, paste the source text and ask about that.
Picture first, question after
Google recommends putting an image before your text, and audio after it.
Start fresh when it drifts
Long chats collect clutter. A new conversation with a short summary often works better.
Check what matters
Like every language model it can state wrong things confidently. Verify facts, figures and code before relying on them.
More techniques in the prompting guide; known weak spots in limitations and safety.
How a Gemma 4 conversation is written
Chat apps and libraries build this for you. It is worth knowing when you write prompts by hand or something looks wrong.
<|turn>system You are a helpful assistant.<turn|> <|turn>user Hello.<turn|> <|turn>model
The prompt ends with an open model turn, which is where Gemma starts writing.
| Piece | Meaning |
|---|---|
<|turn> … <turn|> | Start and end of one turn |
system · user · model | Who is speaking. The reply role is model. |
<|think|> | Placed in the system turn to switch thinking on |
<|channel> … <channel|> | Wraps the model's private reasoning |
From Google's Gemma 4 prompt formatting guide. This differs from earlier Gemma generations, so templates written for Gemma 3 will not work - see the Gemma 3 page. Tool-calling tokens are covered under capabilities.
Drop the thoughts before the next turn
When thinking is on, each reply contains private reasoning and then the answer. Google's guidance is to remove the reasoning from earlier turns before sending the conversation back, so it does not eat context or confuse later replies.
One exception: while the model is in the middle of a tool call, keep its thoughts until that turn is finished.
Build your own chat
Two short starting points: Google's hosted API, and a fully local loop.
# pip install google-genai · set GEMINI_API_KEY first
from google import genai
from google.genai import types
client = genai.Client()
chat = client.chats.create(
model="gemma-4-31b-it",
config=types.GenerateContentConfig(
system_instruction="You are a friendly travel guide.",
),
)
print(chat.send_message("Plan a quiet day in Kyoto.").text)
print(chat.send_message("Swap the evening for something indoors.").text)
The chat object keeps the history for you. Get a key and see limits on the API page.
# pip install openai · ollama pull gemma4:12b-it-qat from openai import OpenAI client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama") history = [{"role": "system", "content": "You are a friendly travel guide."}] while True: history.append({"role": "user", "content": input("You: ")}) reply = client.chat.completions.create( model="gemma4:12b-it-qat", messages=history, ).choices[0].message.content print("Gemma:", reply) # Keep only the final answer in history history.append({"role": "assistant", "content": reply})
Ollama's OpenAI-compatible endpoint uses the role name assistant and converts it to Gemma's model role for you.
Serving many users? See self-hosting and the runtime selector. For chat over your own documents, pair Gemma 4 with EmbeddingGemma 2. For a coding assistant, see the Claude Code guide.
Common chat problems
| What you see | Likely cause | Try this |
|---|---|---|
| Strange tokens or broken formatting | The chat template is missing or out of date | Re-download the model and update your app; the July 2026 refresh fixed the template |
| Reasoning shows up inside the answer | The app is not separating the thought channel | Update the app, or turn thinking off |
| It forgets the start of the chat | The app's context window is smaller than the model's | Raise the context length, for example /set parameter num_ctx 32768 in Ollama |
| Replies are very slow | The model does not fit in memory, or thinking is on | Use a smaller size or a QAT build; turn thinking off for casual chat |
| It ignores your system instruction | The app is pasting it into the user turn | Use an up-to-date app that supports Gemma 4's system role |
| It does not know recent events | Training data ends in January 2025 | Paste the information in, or use a tool with search |
The causes and fixes here are this site's guidance, not an official list. More help on the FAQ page.
Chat questions
What is the quickest way to chat with Gemma 4?
Google AI Studio: sign in, choose a Gemma 4 model and type. There is nothing to install.
Is chatting with Gemma 4 free?
Running it yourself with Ollama, LM Studio or AI Edge Gallery is free - the models are open under Apache 2.0. Google AI Studio is free to try within its limits. Hosting providers charge per token.
Can I chat with it offline?
Yes. Once a model is downloaded to your computer or phone it needs no internet connection.
Can I send pictures or voice?
Every model reads images. Audio works on E2B, E4B and 12B only. All output is text. Details on the features page.
Does Gemma 4 remember past conversations?
No. It only sees what is in the current conversation. Any memory between chats comes from the app you use.
Is this the same as chatting with Gemini?
No. Gemini is Google's hosted product family; Gemma 4 is the open model family you can download and run. See the model comparison.
How good is it at conversation?
On the public Arena text leaderboard, checked October 2, 2026, the 31B scored 1453 and the 26B A4B 1437. More on the benchmarks page and in Gemma 4 vs Qwen 3.5.
Does this website have its own chatbot?
No. This is an independent reference site. It links to official places to chat and does not collect or process chat messages.
Ready to say hello?
Try it in the browser now, or put it on your own machine and keep every word private.