Gemma 4 Prompt Guide

<|turn>system
You are a helpful assistant.<turn|>
<|turn>user
Hello.<turn|>
<|turn>model

The prompt ends with an open model turn; the model writes its answer there. Gemma 4 has a real system role, and its reply role is called model.

<|turn>system
<|think|><turn|>
<|turn>user
What is the water formula?<turn|>
<|turn>model
<|channel>thought
...
<channel|>The most common interpretation of "the water formula" refers...<turn|>

text = processor.apply_chat_template(
    message,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,   # False for fast replies
)

from google import genai
from google.genai import types

client = genai.Client()

response = client.models.generate_content(
    model="gemma-4-26b-a4b-it",
    contents="What is the water formula?",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="high")   # or "minimal"
    ),
)

print(response.text)

# inside an Ollama chat
/set think                       # reason before answering
/set nothink                     # answer straight away

# in the Ollama API, add this field to the request
"think": false

The raw form is from Google's prompt formatting guide: <|think|> at the start of the system turn, with the reasoning returned between <|channel>thought and <channel|>.

• Remove old reasoning. Google: strip the model's thoughts from a turn before sending the conversation back for the next one. Keep only the final answer.
• Except during tool calls. Within a single model turn that calls tools, the thoughts must stay in place between the calls.
• Expect an empty block sometimes. With thinking off, the model may still produce an empty thought block. Strip it like any other.
• You can ask for less. Google notes a system instruction asking for low thinking cut thinking tokens by roughly 20% in its test, and calls this a proof of concept.

Google's audio prompts

# speech to text
Transcribe the following speech segment in {LANGUAGE} into {LANGUAGE} text.

# speech translation
Transcribe the following speech segment in {SOURCE_LANGUAGE}, then translate it into {TARGET_LANGUAGE}.

Wording from the model card. Fill in the language names.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("google/embeddinggemma-2")

# queries and documents get different prompts
query_emb = model.encode(query, prompt_name="SearchQuery")
doc_emb = model.encode(document, prompt_name="Document")

# what the prompts expand to
task: search result | query: how do I reset my router?
title: Router manual | text: Hold the reset button for ten seconds...
title: none | text: A passage that has no title.

# images, audio and video take no task prompt
image_emb = model.encode({"image": "sunset_beach.jpg"})
audio_emb = model.encode({"audio": "ocean_waves.wav"})

# mixed input: placeholders mark where each file belongs
listing_emb = model.encode({
    "text": "Waterproof trail shoe. <|image|> Grip test: <|video|>",
    "image": "trail_shoe.jpg",
    "video": "grip_test.mp4",
})

Code from Google's developer guide; prompt strings from the EmbeddingGemma 2 model card.