Gemini AI

Gemini is Google DeepMind's family of multimodal large language models. It's accessed either through Google AI Studio / the Gemini API (generativelanguage.googleapis.com, simplest path, API-key auth) or through Vertex AI on Google Cloud (enterprise path — IAM auth, VPC controls, regional deployment, the same underlying models).

Model family shape

Google ships Gemini in size/latency tiers within each generation (naming has included, historically, "Pro" for the flagship-capability tier and "Flash" for a faster/cheaper tier, plus lighter "Flash-Lite"/nano-scale variants for on-device or high-volume use). The specific generation number and exact model ids change over time — check Google's model list rather than assuming a name from training data is still current or still maps to the same capability tier.

Core API shape

from google import genai

client = genai.Client(api_key="...")
response = client.models.generate_content(
    model="gemini-2.5-flash",   # verify current model id against docs
    contents="Explain how photosynthesis works in one sentence.",
)
print(response.text)

The core call is generateContent (or streamGenerateContent for streaming). Requests are a list of contents, each with a role (user/model) and one or more parts — which is what makes multimodal input a first-class case rather than a special mode:

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents=[
        {"role": "user", "parts": [
            {"text": "What's in this image?"},
            {"inline_data": {"mime_type": "image/png", "data": image_bytes}},
        ]},
    ],
)

Distinctive capabilities

• Native multimodality — image, audio, video, and PDF input alongside text, across the model family rather than a separate vision-only model.
• Long context windows — Gemini models have generally led on raw context length among major providers; useful for whole-codebase or whole-document workloads. Verify the current window size per model, it varies by tier.
• Function calling / tool use — declare functions with a JSON-schema-like FunctionDeclaration; the model returns a structured call for your code to execute and feed back, the same general pattern as other providers' tool use.
• Context caching — pay a reduced rate for repeatedly reusing a large, unchanged context (e.g. a big document) across multiple calls.
• Grounding with Google Search — an opt-in tool that lets the model retrieve and cite live search results as part of its answer.

Common pitfalls

• Mixing up the Gemini API and Vertex AI auth models — the Gemini API uses a simple API key; Vertex AI uses Google Cloud IAM/service accounts and a project/region. Client libraries and even model ids can differ slightly between the two paths for the same underlying model.
• Assuming a contents/parts shape from an old snippet still matches — the SDKs have gone through more than one major version (google- generativeai → google-genai); check which package a snippet targets.
• Hard-coding a model generation name in code meant to keep working — Google deprecates older generations on a schedule; pin to a specific dated/stable id in production rather than a bare tier name if you need long-term stability.

Learn more

• Gemini API docs
• Google AI Studio — the quickest way to try a prompt and get the equivalent API call.
• Vertex AI Generative AI docs