Skip to content

Google AI Models — Complete Guide

Google’s AI model portfolio divides into two distinct families: Gemini (cloud-hosted, frontier models) and Gemma (open-weight, self-hostable models). Both families support multimodal inputs and share architectural DNA with Google DeepMind’s research.


🗂️ Family🎯 Purpose🏃 Where It Runs📜 Licence
🔵 Gemini 3.xLatest frontier modelsCloud (Gemini API · Google Cloud)Proprietary
🔵 Gemini 2.5Previous frontier generationCloud (Gemini API · Google Cloud)Proprietary
🟢 Gemma 4Open-weight general LLM (2026)Self-hosted · CloudApache 2.0
🟢 Gemma 3Open-weight general LLM (2025)Self-hosted · CloudApache 2.0
🩺 MedGemmaMedical specialised open modelSelf-hostedApache 2.0
🛡️ ShieldGemmaSafety and content filteringSelf-hostedApache 2.0

🤖 Model📅 Released⚙️ Context🎯 Best For⚡ Speed
🚀 Gemini 3.5 FlashMay 20261M tokensAgentic tasks, production speedFastest
💎 Gemini 3.1 ProFeb 20261M tokensResearch, complex coding, analysisSlower
⚖️ Gemini 3 FlashDec 20251M tokensBalanced production workloadsFast
🧠 Gemini 2.5 Pro20251M tokensComplex reasoning, multimodalMedium
⚡ Gemini 2.5 Flash20251M tokensSpeed + intelligence balanceFast
⚡ Gemini 2.5 Flash-Lite20251M tokensHigh-volume, cost-efficientFastest (2.5 gen)

🚀 Gemini 3.5 Flash — Fastest Frontier Model (May 2026)
🔧 Property📋 Detail
📅 ReleasedMay 2026 (Google I/O 2026)
📄 Context window1,000,000 tokens
🎯 FocusAgentic tasks · fast production workloads
🌐 AccessGemini API · Google AI Studio · Google Cloud Vertex AI

🏆 Benchmarks:

  • Beats Gemini 3.1 Pro on 11 of 15 published benchmarks
  • First Flash-tier model to compete with Pro-tier in agentic workflows

✅ Ideal for:

  • 🤖 Agentic pipelines with many tool calls
  • ⚡ High-speed production chat and assistants
  • 💻 Code generation and explanation
  • 📊 Document analysis at scale
  • ⏱️ Latency-sensitive applications needing frontier intelligence

💡 Example:

“Given this product requirements document and codebase, plan the implementation steps, identify risks, and generate the skeleton code with test placeholders.”

💎 Gemini 3.1 Pro — Premium Research and Coding (Feb 2026)
🔧 Property📋 Detail
📅 ReleasedFebruary 2026
📄 Context window1,000,000 tokens
🎯 FocusComplex reasoning · advanced coding · scientific analysis
🌐 AccessGemini API · Google AI Studio · Google Cloud Vertex AI

🏆 Benchmarks:

  • 🔢 80.6% on SWE-bench Verified (software engineering)
  • 🔬 94.3% on GPQA Diamond (scientific reasoning)
  • 🧩 77.1% on ARC-AGI-2

✅ Ideal for:

  • 🔬 Scientific research and academic analysis
  • 🏗️ Complex software architecture decisions
  • 📄 Long-context document analysis (research papers, legal, financial)
  • 🧠 Tasks requiring the deepest available reasoning

💡 Example:

“Analyse this clinical trial dataset (attached), identify statistically significant patterns, and write a structured findings section suitable for peer review.”

⚖️ Gemini 3 Flash — Balanced Default (Dec 2025)
🔧 Property📋 Detail
📅 ReleasedDecember 2025
📄 Context window1,000,000 tokens
🎯 FocusFrontier intelligence at production speed
🌐 AccessGemini API · Google AI Studio · Gemini app (default)

✅ Ideal for:

  • 💬 Production chat assistants
  • 📝 Writing, summarisation, and editing
  • 💻 Everyday coding and explanation
  • 🔄 General-purpose multimodal workloads

🧠 Gemini 2.5 — Previous Generation (Still Supported)

Section titled “🧠 Gemini 2.5 — Previous Generation (Still Supported)”
🧠 Gemini 2.5 Pro — Deep Reasoning
🔧 Property📋 Detail
📄 Context window1,000,000 tokens
🎯 FocusAdaptive thinking · complex agentic + multimodal challenges
🌐 AccessGemini API · Google Cloud Vertex AI

✅ Best for:

  • 🔬 Tasks that need the model to “think before answering” (adaptive thinking mode)
  • 🤖 Complex agentic and multimodal workflows
  • 📄 Very long document or codebase analysis
⚡ Gemini 2.5 Flash — Speed + Intelligence
🔧 Property📋 Detail
📄 Context window1,000,000 tokens
🎯 FocusBalance of intelligence and latency
🌐 AccessGemini API · Google Cloud Vertex AI
🧠 Special featureControllable “thinking budget” — adjust reasoning depth per request

✅ Best for:

  • ⚡ Production workloads needing tunable reasoning depth
  • 💬 Assistants where latency and quality both matter
  • 🔄 Versatile multimodal tasks
⚡ Gemini 2.5 Flash-Lite — High-Volume Efficiency
🔧 Property📋 Detail
📄 Context window1,000,000 tokens
🎯 FocusMassive scale — cost and speed
🌐 AccessGemini API

✅ Best for:

  • 🏷️ Classification and extraction at very high volume
  • 📱 Latency-critical mobile or edge API workloads
  • 💰 Cost-optimised production pipelines

🟢 Gemma — Google’s Open-Weight Models

Section titled “🟢 Gemma — Google’s Open-Weight Models”

The Gemma family gives developers open-weight models they can download, fine-tune, and self-host — built on the same architecture and training approach as Gemini.

🟢 Gemma 4 — Latest Open Model (April 2026)
🔧 Property📋 Detail
📅 ReleasedApril 2, 2026
📜 LicenceApache 2.0 — free to use, modify, and distribute
⚙️ SizesE2B · E4B · 26B A4B · 31B
🌐 AccessGoogle AI · Hugging Face · Kaggle

💡 What “E2B” and “26B A4B” mean:

  • E2B / E4B = 2B and 4B effective parameter MoE models (only a subset of parameters active at inference)
  • 26B A4B = 26B total parameters, 4B active — large-model quality at small-model inference cost
  • 31B = dense model variant

✅ Best for:

  • 🏠 Self-hosted deployment on local servers or cloud VMs
  • 🔧 Fine-tuning on proprietary domain data
  • 🔒 Privacy-sensitive workloads where data cannot leave the organisation
  • 💰 Very high-volume inference where API costs are prohibitive
🟢 Gemma 3 — Production Open Model (March 2025)
🔧 Property📋 Detail
📅 ReleasedMarch 12, 2025
📜 LicenceApache 2.0
⚙️ Sizes1B · 4B · 12B · 27B
📄 Context window128,000 tokens
👁️ ModalityText + image input (all sizes)
🌍 Languages140+ languages
🌐 AccessGoogle AI · Hugging Face · Kaggle · NVIDIA NIM

✅ Best for:

  • 📱 Mobile and edge deployment (1B, 4B sizes)
  • 💬 Self-hosted chat assistants
  • 🔍 RAG pipelines on private data
  • 💻 Code generation and explanation (12B, 27B sizes)
  • 🌍 Multilingual applications

📦 Gemma 3n — Mobile-Optimised Variant:

Gemma 3n is specifically optimised for phones, laptops, and tablets — designed for on-device AI without a GPU.


🩺 MedGemma — Medical AI (4B and 27B)
🔧 Property📋 Detail
⚙️ Sizes4B · 27B
🎯 FocusMedical image analysis and clinical text understanding
📜 LicenceApache 2.0

✅ Use for:

  • 🏥 Medical image analysis (radiology, pathology)
  • 🩺 Clinical text understanding and extraction
  • 🔬 Biomedical research assistance

⚠️ Not a replacement for clinical judgement — requires validation for any medical decision support use.

🛡️ ShieldGemma 2 — Safety and Content Filtering (4B)
🔧 Property📋 Detail
⚙️ Size4B
🎯 FocusClassifying violent, dangerous, or sexually explicit images
📜 LicenceApache 2.0

✅ Use for:

  • 🛡️ User-generated content moderation
  • 🚨 Abuse detection pipelines
  • 🔒 Pre-screening images before passing to other models

🎯 Task✅ Best Model💡 Why
🤖 Agentic pipeline with many tool callsGemini 3.5 FlashBest agentic performance at Flash speed
🔬 Research or complex scientific analysisGemini 3.1 ProHighest reasoning depth
💻 Complex coding or architecture decisionsGemini 3.1 ProStrong SWE-bench scores
💬 Production chat assistantGemini 3 FlashBalanced frontier intelligence at speed
🧠 Task needing tunable reasoning depthGemini 2.5 FlashControllable thinking budget
🏷️ High-volume classification/extractionGemini 2.5 Flash-LiteCheapest in the 2.5 family
🏠 Self-hosted / on-premiseGemma 4 (26B A4B)Best open-weight quality
📱 Mobile or edge device deploymentGemma 3 (1B/4B) or Gemma 3nSmallest and most efficient
🩺 Medical image or clinical textMedGemma (27B)Specialist medical fine-tune
🛡️ Image safety filteringShieldGemma 2 (4B)Built for content classification

🏷️ Platform🤖 Models Available
🔵 Google AI Studio (ai.google.dev)All Gemini models · Free tier available
🔵 Gemini APIAll Gemini models · Pay-per-token
☁️ Google Cloud Vertex AIGemini models · Enterprise SLA
💬 Gemini AppGemini 3 Flash (default) · Pro via Gemini Advanced
🤗 Hugging FaceFull Gemma family
🐙 KaggleFull Gemma family
🟩 NVIDIA NIMGemma 3 family