Google DeepMind and Gemini

Models and offerings #

Who they are

This hub is for people choosing a hosted model, people who want an open-weight model with a clean licence, and people who have TPU credits. Google Brain published the transformer paper in 2017 and BERT in 2018, and DeepMind (acquired 2014, London) published AlphaGo and AlphaFold. The two were merged into Google DeepMind in 2023 and the first Gemini models shipped in December that year. Gemma, the open-weight line, followed in February 2024.

Google’s position is unusual in owning every layer: the research lab, the models, the serving product (Gemini API and Vertex AI), the framework (JAX) and the chips (TPU). It is also the only large lab to have moved its open models to a permissive licence: Gemma 4, released in April 2026, is Apache-2.0, where Gemma 1 to 3n used custom terms.

Tooling and APIs

Two front doors. AI Studio at aistudio.google.com gives a browser playground and an API key in a minute, and pip install google-genai installs the SDK (the older google-generativeai package is deprecated). Vertex AI is the Google Cloud route with IAM, data residency and a model garden that also sells Claude and open models. Models run only on Google’s infrastructure, while Gemma runs anywhere PyTorch, JAX, llama.cpp or Ollama runs.

# needs-api
import os
from google import genai

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
response = client.models.generate_content(
    model="gemini-4-hydrogen",
    contents="In one sentence, what does a learning rate control?",
)
print(response.text)

Prices at the time of writing, per million tokens, from the Gemini API pricing page: Gemini 4 Carbon (described by Google as its most capable reasoning model) and Gemini 4 Argon are 2 USD input and 10 USD output, rising to 4 and 20 on 1 December 2026. Gemini 4 Hydrogen is 1.50 and 7.50 with a free tier. Gemini 3.8 Flash is 0.75 and 3.75, doubling on 1 January 2027, and Gemini 2.5 Flash-Lite is 0.10 and 0.40. All Gemini 4 models have a 1 M token context window and 65 k output. Release dates and parameter counts are not stated on the model pages. The pricing page also records that requests to Gemini 2.5 Pro will be rerouted to Gemini 3.8 Flash on 1 December 2026, which is the clearest example of why you pin model ids.

Gemma 4 comes in five sizes according to the model card: E2B (5.1 B total with embeddings, 2.3 B effective, 128 k context, text, image and audio), E4B (8 B total, 4.5 B effective, 128 k), 12 B (256 k), 26B A4B (25.2 B total, 3.8 B active, 256 k) and 31 B dense (256 k). All are Apache-2.0 and on Hugging Face and Kaggle. Pain points: the Gemini free tier has low rate limits, and according to Google’s API terms prompts sent on the unpaid tier may be used to improve its products while paid-tier prompts are not. Vertex AI and AI Studio have different auth, quotas and sometimes prices for the same model.

Migration notes

From OpenAI or Anthropic: the google-genai SDK has its own shape (contents, config, GenerateContentConfig), with the system prompt as system_instruction. The Gemini API also exposes an OpenAI-compatible endpoint for chat completions, which covers simple cases but not every feature. Multimodal input is where Gemini differs most: audio, video and PDF are first-class content parts with per-second or per-page pricing rather than image-only.

Gemma to and from other open models: Gemma 4 weights are standard safetensors, so Transformers, vLLM, llama.cpp (GGUF) and MLX conversions all exist. The tokeniser differs from Llama’s, so adapters do not transfer. Earlier Gemma checkpoints carry the Gemma Terms of Use, which treat hosted access as distribution and include a use policy, so a Gemma 3 fine-tune does not become Apache-2.0 because Gemma 4 is.

Training on TPU means JAX or PyTorch/XLA. See the JAX hub and the frameworks reference.

Watch for

Scheduled price rises are printed on the pricing page months ahead, which is better than most vendors, but they are rises: the Gemini 4 and 3.8 Flash prices double on stated dates. Automatic rerouting of retired models to a different model changes behaviour without an error.

Free-tier data use is the main privacy sharp edge, and the consumer Gemini app has its own settings. Vertex AI pricing can differ from AI Studio for the same model id and region. Quotas are per project and per model, and bursting past them returns 429s that the SDK retries silently, which looks like latency.

Gemma 4 is open weights with a permissive licence, but the training data and code are not published, so it is not open source in the OSI sense. Benchmark figures on the Gemma and Gemini pages are the vendor’s reports. Google retires model versions on a schedule and deprecated ids stop working, so track the deprecation table rather than discovering it in production.