API access to Google's Gemini models
Install this plugin in the same environment as LLM.
llm install llm-geminiConfigure the model by setting a key called "gemini" to your API key:
llm keys set gemini<paste key here>
You can also set the API key by assigning it to the environment variable LLM_GEMINI_KEY.
Now run the model using -m gemini-flash-latest, for example:
llm -m gemini-flash-latest "A short joke about a pelican and a walrus"A pelican and a walrus are sitting at a bar. The pelican orders a fishbowl cocktail, and the walrus orders a plate of clams. The bartender asks, "So, what brings you two together?"
The walrus sighs and says, "It's a long story. Let's just say we met through a mutual friend... of the fin."
You can set the default model to avoid the extra -m option:
llm models default gemini-flash-latest
llm "A joke about a pelican and a walrus"gemini/gemini-3.5-flash-lite: Gemini 3.5 Flash Litegemini/gemini-3.6-flash: Gemini 3.6 Flashgemini/gemini-3.5-flash: Gemini 3.5 Flashgemini/gemini-3.1-flash-lite: Gemini 3.1 Flash Litegemini/gemma-4-31b-it: Gemma 4 31B Instructgemini/gemma-4-26b-a4b-it: Gemma 4 26B-A4B Instructgemini/gemini-3.1-flash-lite-preview: Gemini 3.1 Flash Lite Previewgemini/gemini-3.1-pro-preview-customtoolsgemini/gemini-3.1-pro-preview: Gemini 3.1 Pro Previewgemini/gemini-3-flash-preview: Gemini 3 Flash Previewgemini/gemini-flash-lite-latest: Latest Gemini Flash Litegemini/gemini-flash-latest: Latest Gemini Flashgemini/gemini-2.5-flash: Gemini 2.5 Flash
All of these models have aliases that omit the gemini/ prefix, for example:
llm -m gemini-flash-latest --schema 'name,age int,bio' 'invent a dog'Gemini models are multi-modal. You can provide images, audio or video files as input like this:
llm -m gemini-flash-latest 'extract text' -a image.jpgOr with a URL:
llm -m gemini-flash-latest 'describe image' \
-a https://static.simonwillison.net/static/2024/pelicans.jpgAudio works too:
llm -m gemini-flash-latest 'transcribe audio' -a audio.mp3And video:
llm -m gemini-flash-latest 'describe what happens' -a video.mp4The Gemini prompting guide includes extensive advice on multi-modal prompting.
You can provide YouTube video URLs as attachments as well:
llm -m gemini-flash-latest -a 'https://www.youtube.com/watch?v=9o1_DL9uNlM' \
'Produce a summary with relevant URLs and code example snippets, then an accurate transcript with timestamps.'These will be processed with media resolution low by default. You can use the -o media_resolution X option to set that to medium, high, or unspecified.
Use -o json_object 1 to force the output to be JSON:
llm -m gemini-flash-latest -o json_object 1 \
'3 largest cities in California, list of {"name": "..."}'Outputs:
{"cities": [{"name": "Los Angeles"}, {"name": "San Diego"}, {"name": "San Jose"}]}Gemini models can write and execute code - they can decide to write Python code, execute it in a secure sandbox and use the result as part of their response.
Enable this server-side tool with -T CodeExecution:
llm -m gemini-3.6-flash -T CodeExecution \
'use python to calculate (factorial of 13) * 3'Some Gemini models support Grounding with Google Search, where the model can run a Google search and use the results as part of answering a prompt.
Using this feature may incur additional requirements in terms of how you use the results. Consult Google's documentation for more details.
Enable this server-side tool with -T GoogleSearch:
llm -m gemini-3.6-flash -T GoogleSearch \
'What happened in Ireland today?'The plugin leaves the model's response text unchanged and retains Gemini's raw
groundingMetadata on the response part. Use llm logs -c --json after running
a prompt to inspect that metadata, which includes additional information about grounded results.
When Gemini returns native server-side tool invocation parts, the plugin exposes those as structured server-side tool call and result events as well.
Gemini models support a URL context tool which, when enabled, allows the models to fetch additional content from URLs as part of their execution.
Enable this server-side tool with -T URLContext - for example:
llm -m gemini-2.5-flash -T URLContext 'Latest headline on simonwillison.net'Extra tokens introduced by this tool will be charged as input tokens. Use --usage to see details of those:
llm -m gemini-2.5-flash -T URLContext --usage \
'Latest headline on simonwillison.net'Outputs:
The latest headline on simonwillison.net as of August 17, 2025, is "TIL: Running a gpt-oss eval suite against LM Studio on a Mac.".
Token usage: 9,613 input, 87 output, {"candidatesTokenCount": 57, "promptTokensDetails": [{"modality": "TEXT", "tokenCount": 10}], "toolUsePromptTokenCount": 9603, "toolUsePromptTokensDetails": [{"modality": "TEXT", "tokenCount": 9603}], "thoughtsTokenCount": 30}
The "toolUsePromptTokenCount" key shows how many tokens were used for that URL context.
To chat interactively with the model, run llm chat:
llm chat -m gemini-flash-latestBy default there is no timeout against the Gemini API. You can use the timeout option to protect against API requests that hang indefinitely.
With the CLI tool that looks like this, to set a 1.5 second timeout:
llm -m gemini-flash-latest 'epic saga about mice' -o timeout 1.5In the Python library timeouts are used like this:
import httpx, llm
model = llm.get_model("gemini/gemini-flash-latest")
try:
response = model.prompt(
"epic saga about mice", timeout=1.5
)
print(response.text())
except httpx.TimeoutException:
print("Timeout exceeded")An httpx.TimeoutException subclass will be raised if the timeout is exceeded.
The plugin supports Google's current Gemini embedding models:
gemini-embedding-2is the latest model.gemini-embedding-001remains available for text-only use cases.
Run that against a single string like this:
llm embed -m gemini-embedding-2 -c 'hello world'This returns a JSON array of 3072 numbers.
Both models have variants that ask Gemini to return its recommended smaller vector sizes of 768 or 1536 dimensions, specified as a suffix on the model ID:
gemini-embedding-2- 3072 numbersgemini-embedding-2-1536- 1536 numbersgemini-embedding-2-768- 768 numbersgemini-embedding-001- 3072 numbersgemini-embedding-001-1536- 1536 numbersgemini-embedding-001-768- 768 numbers
The embedding spaces used by the two models are incompatible. If you switch an
existing collection from gemini-embedding-001 to gemini-embedding-2, you
must re-embed all of its content.
This command will embed every README.md file in child directories of the current directory and store the results in a SQLite database called embed.db in a collection called readmes:
llm embed-multi readmes -d embed.db -m gemini-embedding-2-768 \
--files . '*/README.md'You can then run similarity searches against that collection like this:
llm similar readmes -c 'upload csvs to stuff' -d embed.dbSee the LLM embeddings documentation for further details.
The llm gemini models command lists all of the models that are exposed by the Gemini API, some of which may not be available through this plugin.
llm gemini modelsYou can add a --key X option to use a different API key.
To filter models by their supported generation methods use --method one or more times:
llm gemini models --method embedContentIf you provide multiple methods you will see models that support any of them.
To set up this plugin locally, first checkout the code, then run the tests with uv:
cd llm-gemini
uv run pytestRun llm with the plugin like this:
uv run llm models -q geminiThis project uses pytest-recording to record Gemini API responses for the tests.
If you add a new test that calls the API you can capture the API response like this:
PYTEST_GEMINI_API_KEY="$(llm keys get gemini)" uv run pytest --record-mode onceYou will need to have stored a valid Gemini API key using this command first:
llm keys set gemini
# Paste key here