The Gemini Family (Google), with PaLM, Bard and Gemma¶

In one line: Gemini is Google's closed, natively multimodal model line that replaced PaLM and absorbed the Bard brand, sold as Pro / Flash / Flash-Lite tiers with a frequent-update Flash line, while Gemma is its smaller open-weight sibling - read a name as "generation, then tier". Last reviewed: 2026-09-30

The short version¶

• Two eras at Google. First came separate research models (LaMDA for dialogue, PaLM for scale). In 2023 Google merged Google Brain and DeepMind into Google DeepMind, and Gemini (December 2023) became the single model family behind Google's products.
• Multimodal and long-context from the start. Gemini was trained on text, images, audio and video together, not as a text model with add-ons. Gemini 1.5 (February 2024) made million-token context a headline feature.
• Mixture-of-experts at the core. Gemini 1.5 introduced a mixture-of-experts (MoE) design, and Google's Gemini 3 Pro model card describes it as a sparse MoE transformer. MoE means only part of the network runs for each token, so total size and per-token cost are decoupled.
• Flash became the fast-moving line. Since Gemini 3.5 (May 2026) Google has shipped a new Flash roughly monthly (3.5, 3.6, 3.7, 3.8), while the Pro tier has stayed at 3.1 Pro since February 2026. Gemini 4 is in post-training as of September 2026.
• Gemini is closed; Gemma is open. Gemma models (from February 2024) are released as downloadable weights. Gemma 4 (March 31, 2026) moved to the Apache 2.0 licence.

Release timeline¶

"Summary" links go to paper summaries in this repo where one exists. Image, speech and robotics offshoots are left out except where they mark a turn in the family.

Before Gemini¶

Gemini¶

Gemma (open weights)¶

Around the core line Google has also released specialised Gemma variants, for example CodeGemma, PaliGemma (vision), ShieldGemma (safety classification), MedGemma (medical), EmbeddingGemma and VaultGemma (trained with differential privacy).

The through-line¶

1. Scale, then consolidate¶

Google's research labs produced many of the ideas every model family now uses: the transformer itself, BERT, T5, mixture-of-experts at scale, chain-of-thought prompting and Chinchilla's compute-optimal scaling. But until 2023 those ideas were spread across separate model lines. PaLM showed Google could train at the largest scale; ChatGPT showed that product was what mattered. Gemini is the consolidation: one family, one lab (Google DeepMind), deployed across Search, Workspace, Android and Cloud.

2. Native multimodality¶

Where other families added vision to a text model, Gemini was trained from the start on interleaved text, images, audio and video. That choice is why Gemini has tended to lead on video and long-document understanding, and why Google ships voice ("Live"), image ("Nano Banana") and robotics variants from the same base. See CLIP and LLaVA for the bolt-on approach it is contrasted with.

3. Long context as a feature¶

Gemini 1.5 Pro's 1M-token context (February 2024) was the first time a major model could take in hours of video or a whole codebase in one prompt. The 1M-input, 64K-output shape has held from Gemini 3 Pro through Gemini 3.8 Flash. A big context window is not the same as using it well; see context windows and the RULER benchmark for how effective context is measured.

4. Sparse compute: MoE plus distillation¶

            Pro (largest, slowest to update)
             |   knowledge passed down by distillation
             v
            Flash (workhorse, updated roughly monthly in 2026)
             |
             v
            Flash-Lite (cheapest, highest throughput)

   each: sparse mixture-of-experts -> only some "experts" run per token

Sparse MoE lets Google grow total capacity without growing per-token cost in proportion. The tier structure leans on knowledge distillation, where a smaller model is trained to imitate a larger one. See mixture of experts for the architecture, and the hands-on quantization and distillation page.

5. Thinking and agents¶

Gemini 2.5 (March 2025) made "thinking" the default across the family, with a thinking budget developers can set. Deep Think adds a heavier reasoning mode for the hardest problems. By 2026 launch posts lead with agentic coding and tool use: Gemini 3.8 Flash's own announcement says its gains come from executing extra reasoning steps and calling tools iteratively. Background: reasoning models and test-time compute.

6. Security-specialised variants¶

In 2026 Google began shipping "Cyber" variants of Flash (3.5 Flash Cyber in July, 3.8 Flash Cyber in September) aimed at vulnerability discovery. Both are restricted: 3.5 Flash Cyber to governments and trusted partners, 3.8 Flash Cyber to government, critical-infrastructure and software-maintainer applicants through Google's Fairwind Program. The general models ship with safeguards against chemical, biological, radiological and nuclear (CBRN) misuse and cyber offence. This mirrors the gated-access pattern in the Claude and GPT families; see frontier safety frameworks.

Open vs closed¶

Google runs a two-track strategy: its best models stay closed, while Gemma gives developers something to download, fine-tune and run on their own hardware. The switch to Apache 2.0 for Gemma 4 removed custom licence terms, putting Gemma on the same footing as other permissively licensed open models such as Qwen. See open vs closed weights and the open-source stack.

How to read the naming¶

Gemini  3.8   Flash   (-Lite / Cyber / Live / Image ...)
          |     |              |
          |     |              +-- specialised variant or modality
          |     +-- tier: Pro > Flash > Flash-Lite   (Ultra and Nano in 1.0 only)
          +-- generation.point release

• Tiers. Gemini 1.0 had Ultra, Pro and Nano. From 1.5 the main tiers became Pro and Flash, with Flash-Lite added in 2.0. "Ultra" now mostly names Google's top consumer subscription (Google AI Ultra), not a model.
• Point numbers move per tier. As of September 30, 2026 the latest Pro is 3.1, the latest Flash is 3.8, and the latest Flash-Lite is 3.5. A higher number in one tier does not imply a matching release in another. Google's July 2026 post said Gemini 3.5 Pro was "testing with partners".
• Deep Think is a reasoning mode on top of a model, not a separate generation.
• Nano Banana is the public nickname for Gemini's image generation and editing models (for example Gemini 2.5 Flash Image), not a language model tier.
• Preview, then stable. API model IDs often start with a -preview suffix (for example gemini-3-pro-preview) before a stable ID; aliases like gemini-flash-latest move to the newest stable model.
• Gemma sizes. "E4B" means "effective 4 billion" parameters in memory at run time. "26B A4B" means 26 billion total, about 4 billion active per token (a mixture-of-experts model).

What to watch¶

• Gemini 4. Google said in July 2026 it had started "our most ambitious pre-training run yet, for Gemini 4". In late September 2026 Google DeepMind's Koray Kavukcuoglu said the model was in post-training and that Google hoped to release it "much earlier" than the end of 2026. No date has been announced as of this review.
• The Pro gap. Gemini 3.1 Pro (February 2026) is now seven months old while Flash has had four updates. Whether 3.5 Pro ships or Google goes straight to Gemini 4 will show whether the frequent-Flash, slower-Pro cadence is deliberate.
• Retirements. Gemini 2.0 Flash and Flash-Lite were shut down on June 1, 2026, and from September 18, 2026 access to the 2.5 models is limited to projects with prior usage. Expect the same cycle for early Gemini 3 models.
• Gemma after Apache 2.0. Gemma 4's licence change makes it a stronger base for fine-tuning. Watch whether Gemma keeps pace with the closed line and with open models from Qwen, DeepSeek and Mistral.

Read next¶

• Paper summaries: PaLM, Gemini 2.5, Gemini 3, T5, FLAN, Chinchilla, Switch Transformer, Imagen, RT-2, AlphaFold 3
• Other families: GPT and o-series, Claude, Llama
• Explainers: Context windows, Open vs closed weights, AI hardware landscape (Google trains Gemini on its own TPUs), Labs landscape
• Hands-on: Transformer architecture, GPUs for AI

Sources¶

• Google AI for Developers, Gemini API changelog (model launch and retirement dates, November 2025 to September 2026): https://ai.google.dev/gemini-api/docs/changelog
• Google AI for Developers, Gemma releases (all Gemma dates and sizes): https://ai.google.dev/gemma/docs/releases
• Google, "Gemma 4: Byte for byte, the most capable open models" (Apache 2.0, sizes, 256K context): https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
• Google, "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber": https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/
• Google DeepMind, Gemini 3.8 Flash model card (1M input, 64K output): https://deepmind.google/models/model-cards/gemini-3-8-flash/
• Google, "Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber" (July 21, 2026; 3.5 Pro testing; Gemini 4 pre-training): https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/
• Google, "Gemini 3.1 Pro: A smarter model for your most complex tasks" (February 19, 2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/
• Google DeepMind, Gemini 3 Pro model card (sparse MoE, 1M input, 64K output): https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf
• Google, "Gemini 3 Deep Think is now available in the Gemini app": https://blog.google/products/gemini/gemini-3-deep-think/
• Google, "Introducing Gemini 3 Flash": https://blog.google/products-and-platforms/products/gemini/gemini-3-flash/
• Google, "Bard is now Gemini" (February 8, 2024): https://blog.google/products-and-platforms/products/gemini/bard-gemini-advanced-app/
• Google, "Introducing Gemini: our largest and most capable AI model" (December 6, 2023): https://blog.google/innovation-and-ai/technology/ai/google-gemini-ai/
• Gemini Team, "Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context" (MoE, 1M context): https://arxiv.org/abs/2403.05530
• Gemini Team, "Gemini 2.5" technical report: https://arxiv.org/abs/2507.06261
• 9to5Google, "Google says Gemini 4 release is coming 'as soon as possible'" (September 24, 2026, Kavukcuoglu remarks): https://9to5google.com/2026/09/24/google-says-gemini-4-release-is-coming-as-soon-as-possible/
• Google, "Introducing PaLM 2" (May 10, 2023): https://blog.google/technology/ai/google-palm-2-ai-large-language-model/
• Wikipedia, "Gemini (language model)" (cross-check of release dates): https://en.wikipedia.org/wiki/Gemini_(language_model)