Gemini Family: Multimodal-First from Google

Google DeepMind's Gemini models are built around the assumption that text, images, audio, and video should share a single Transformer rather than being bolted together post hoc. This shows up in tokenization, training data, and the API surface.

Model	Context	Output max	Input/Output ($/1M)	Notable
Gemini 1.5 Pro	1M (then 2M experimental)	8K	Variable	Set the modern long-context bar
Gemini 2.0 Flash	1M	8K	$0.10 / $0.40	Fast, multimodal, very cheap
Gemini 2.0 Pro Experimental	2M	—	—	2M context push
Gemini 2.5 Flash	1M	8K	$0.30 / $2.50	~110 tokens/s output, thinking mode
Gemini 2.5 Pro	1M	8K	$1.25 / $10.00	Flagship, thinking mode, frontier benchmarks

Benchmarks (Gemini 2.5 Pro): AIME 2024 92%, AIME 2025 83%, GPQA 83%, SWE-bench 63%. The Pro series supports a "thinking" mode that allocates extra inference compute to reasoning, comparable to OpenAI's o-series and DeepSeek-R1.

Multimodal: native text + image + audio + video input. Output remains text (with text + image in some endpoints). The single-stack design means Gemini can directly attend across modalities — an image patch and a text token live in the same residual stream.

Code

Calling Gemini with vision·python

import google.generativeai as genai

genai.configure(api_key="...")
model = genai.GenerativeModel("gemini-2.5-pro")

with open("photo.jpg", "rb") as f:
    image_bytes = f.read()

response = model.generate_content([
    {"mime_type": "image/jpeg", "data": image_bytes},
    "Describe what's in this image and what the people seem to be doing.",
])
print(response.text)
# Multimodal in a single API call — image and text share the residual stream.

Gemini Family: Multimodal-First from Google

Code

External links

Exercise

Progress

Comments 0