Premium museum expansion · Provider catalog
Google Gemini
Google DeepMind multimodal family — successors to LaMDA and PaLM 2 in the narrative of Google’s chat and frontier models. Covers Bard through Gemini 3.x, the Gemma open line, and the “Nano Banana” image stack. Synthesized from Gemini (language model), Google Gemini, and related articles (March 2026 snapshot); product names change faster than encyclopedia updates.
Architecture & training (overview)
Gemini is natively multimodal: context can interleave text, images, video, and audio. Video is ingested as image sequences; audio is sampled at 16 kHz and tokenized via the Universal Speech Model. Training data is multimodal and multilingual (web, books, code, images, audio, video). Gemini and Gemma are decoder-only transformers with TPU-oriented training and inference optimizations; Gemini 1.0 used multi-query attention. No dedicated whitepapers were published for Gemini 2.0, 2.5, and 3 in the same way as 1.0 — rely on blog posts, model cards, and press coverage for those generations.
Infrastructure: Google Tensor Processing Units (TPUs). The codename Gemini nods to the DeepMind / Google Brain merger and NASA’s Project Gemini. February 2024: Bard and Duet AI were unified under the Gemini brand; Gemini Advanced (Ultra 1.0) shipped with Google One AI Premium.
Official technology hub: deepmind.google/technologies/gemini/.
Tiers & naming (family overview)
Across generations, Google has used consistent capability bands (exact SKUs vary by year):
- Ultra — largest / most capable; Pro — wide task range; Nano — on-device; Flash — speed; Flash Lite — cost-efficiency; Deep Think — heavy reasoning mode.
- Competitive framing at launch included OpenAI GPT-4 and assistive coding tools in the GitHub Copilot class; training sources described in public reporting include large-scale web data and transcripts of YouTube videos (among many other inputs).
Launch & benchmarks (Gemini 1.0 era)
Announced at Google I/O (May 2023); December 6, 2023 preview with Sundar Pichai and Demis Hassabis; broader rollout including February 8, 2024 in many sources. Gemini Ultra was cited for first-in-class language-model performance on benchmarks such as MMLU (reported 90% on the Massive Multitask Language Understanding suite in launch materials). Gemini Pro was positioned above GPT-3.5-class models in Google’s comparisons at the time.
Google discussed safety and governance with U.S. and U.K. government contexts aligned with executive orders and the Bletchley Park AI Safety Summit narrative. Key figures in public coverage often include Sergey Brin alongside DeepMind and Google leadership.
IDE model menus (Cursor, Antigravity & OpenCode)
The same Gemini 3.1 Pro (High) / (Low) and Gemini 3 Flash labels that appear in Cursor-style pickers also show up in Google Antigravity (Google’s agentic IDE). See the Cursor / Antigravity snapshot on the main AI Model Museum page for Q1 2026 menu wording. OpenCode Desktop uses a different layout: a free-model row (e.g. Big Pickle, MiMo V2, GPT-5 Nano, Nemotron 3 Super, MiniMax M2.5) plus provider add-ons — see OpenCode “Select model” snapshot.
Model versions (consumer / product line)
Chronological product names, release windows, lifecycle status, and short descriptions — aligned with Wikipedia’s “Model versions” style table.
| Version | Release date | Status | Description |
|---|---|---|---|
| Bard | 21 Mar 2023 | Discontinued | First experimental chatbot built on LaMDA; superseded by the Gemini brand. |
| Gemini 1.0 Nano | 6 Dec 2023 | Discontinued | On-device stack; early Pixel 8 Pro deployment. |
| Gemini 1.0 Pro | 13 Dec 2023 | Discontinued | General-purpose tier; Vertex / AI Studio path. |
| Gemini 1.0 Ultra | 8 Feb 2024 | Discontinued | Flagship of the 1.0 series; benchmarked vs GPT-4 class models at launch. |
| Gemini 1.5 Pro | 15 Feb 2024 | Discontinued | Long-context successor (up to ~1M tokens in product); MoE architecture. |
| Gemini 1.5 Flash | 14 May 2024 | Discontinued | Renamed from “Nano” positioning; fast / free-tier workhorse. |
| Gemini 2.0 Flash | 30 Jan 2025 | GA | Multimodal + agentic focus; default in many surfaces after 1.5 Flash. |
| Gemini 2.0 Flash-Lite | 1 Feb 2025 | GA | First Flash-Lite variant — cost and latency optimized. |
| Gemini 2.0 Pro | 5 Feb 2025 | GA | Higher-quality 2.0 tier alongside Flash. |
| Gemini 2.5 Pro | 25 Mar 2025 | GA | Strong reasoning / coding push; “Deep Think” modes appear in this era. |
| Gemini 2.5 Flash | 17 Apr 2025 | GA | Incremental upgrade over 2.0 Flash; default in many UIs. |
| Gemini 2.5 Flash-Lite | 17 Jun 2025 | GA | Enterprise-scale efficiency tier. |
| Gemini 2.5 Flash Image (Nano Banana) | 26 Aug 2025 | GA | Image generation / editing codename Nano Banana; viral “figurine” trend. |
| Gemini 3 Pro | 18 Nov 2025 | Discontinued preview | Sparse MoE; up to ~64K output tokens in materials cited on Wikipedia. |
| Gemini 3 Pro Image (Nano Banana Pro) | 20 Nov 2025 | Preview | Improved text-in-image and real-world knowledge vs first Nano Banana. |
| Gemini 3 Deep Think | 3 Dec 2025 | Preview | Heavy reasoning mode (IOI / olympiad-class narratives in press). |
| Gemini 3 Flash | 17 Dec 2025 | Preview | Speed-first Gemini 3 tier. |
| Gemini 3.1 Pro | 19 Feb 2026 | Preview | Next-gen Pro; “smarter for complex tasks” in Google messaging. |
| Gemini 3.1 Flash Image (Nano Banana 2) | 26 Feb 2026 | Preview | Successor image model; integrated across Gemini, Search AI Mode, Lens per timeline sources. |
| Gemini 3.1 Flash Lite | 3 Mar 2026 | Preview | Intelligence-at-scale / API-forward Lite tier (enterprise narratives). |
See also: open-weight Gemma family (below). Gemini 3 Pro article on Wikipedia may use slightly different dates vs regional rollouts; treat as indicative.
Technical specifications (published 1.x era)
From the Gemini technical report era — parameters and context where disclosed. Later generations often ship without matching public parameter tables.
| Generation | Variant | Release | Parameters | Context length | Notes |
|---|---|---|---|---|---|
| 1.0 | Nano-1 | 6 Dec 2023 | 1.8B | 32,768 | Distilled; 4-bit quantized for on-device. |
| 1.0 | Nano-2 | 6 Dec 2023 | 3.25B | 32,768 | Distilled; 4-bit quantized. |
| 1.0 | Pro | 13 Dec 2023 | ? | 32,768 | Cloud / API tier. |
| 1.0 | Ultra | 8 Feb 2024 | ? | 32,768 | Flagship 1.0. |
| 1.5 | Pro | 15 Feb 2024 | ? | 10,000,000 (test) | ~1M tokens in production API; up to 10M in research claims. |
| 1.5 | Flash / “Mini” | 14 May 2024 | ? | 10,000,000 (test) | Fast tier; consumer default in many flows. |
Nano Banana — image & editing codenames
Google DeepMind’s viral image stack, often discussed under fruit codenames in blogs and Wikipedia:
- Nano Banana → retail name Gemini 2.5 Flash Image (Aug 2025+). Subject consistency, multi-image fusion, natural-language edits, SynthID watermarking.
- Nano Banana Pro → Gemini 3 Pro Image (Nov 2025) — better text rendering and world knowledge.
- Nano Banana 2 → Gemini 3.1 Flash Image (Feb 2026) — faster, stronger instruction following and typography in images.
Gemma (open-weights family)
Gemma is Google DeepMind’s open / terms-based LLM line built with the same research stack as Gemini — often framed as Google’s answer to Meta’s Llama-class openness. License: Gemma License (not Apache; read terms). Model hub: deepmind.google/models/gemma/. English Wikipedia: Gemma (language model).
Gemma 3 (stable release March 12, 2025 in public sources) adds multimodal inputs (text + image), long context (up to 128K tokens on most sizes; 32K on the 1B), 140+ languages, grouped-query attention (GQA), SigLIP vision encoder, function calling, and Gemma 3n variants tuned for phones and laptops. Reported scale: 150M+ downloads and 70K+ Hugging Face derivatives (order-of-magnitude press figures).
Technical specifications (Gemma models)
| Generation | Release | Parameters | Context | Multimodal | Notes |
|---|---|---|---|---|---|
| Gemma 1 | 21 Feb 2024 | 2B, 7B | 8,192 | No | 2B distilled from 7B; 2B uses multi-query attention, 7B multi-head. |
| CodeGemma | 2024 | 2B, 7B | 8,192 | No | Gemma 1–family finetune for code (Python, Java, C++, etc.). |
| RecurrentGemma | 11 Apr 2024 | 2B, 9B | Unlimited (trained 8,192) | No | Griffin recurrent architecture — not a vanilla Transformer. |
| Gemma 2 | 27 Jun 2024 | 2B, 9B, 27B | 8,192 | No | Grouped-query attention; distillation chain from larger teachers. |
| PaliGemma | 10 Jul 2024 | 3B | 8,192 | Image | SigLIP-So400m + Gemma 1 2B; vision-language. |
| PaliGemma 2 | 4 Dec 2024 | 3B, 10B, 28B | 8,192 | Image | SigLIP + Gemma 2 sizes; broader VLM tasks. |
| Gemma 3 | 12 Mar 2025 | 1B, 4B, 12B, 27B | 131,072 | Image | 1B is text-only; local + global attention for long context efficiency. |
Context scaling: early Gemma / PaliGemma lines grow KV-cache roughly linearly with sequence length; Gemma 3 improves the curve via local vs global attention. RecurrentGemma caps memory growth after ~2,048 tokens in public descriptions.
Specialized Gemma-based variants
- ShieldGemma 2 (4B) — safety / moderation (violence, sexual content, etc.).
- MedGemma (4B & 27B) — medical imaging and clinical-assist workflows; not clinical-grade by Google’s disclaimers.
- DolphinGemma (~400M) — marine biology / audio research with Georgia Tech & Wild Dolphin Project; limited public release narrative.
- CodeGemma — see table; coding rival framing vs GitHub Copilot in press.
Selected release timeline (2024–2026)
- Sep 2024: Gemini 1.5 Pro-002 & Flash-002 refreshes.
- Dec 2024: Gemini 2.0 Flash Experimental — Live API (audio/video), native image + TTS, Search grounding; Jules coding agent experimental.
- Feb 2025: Gemini 2.0 Flash Thinking Experimental — optional reasoning summaries.
- Mar 2025: Gemini Robotics (VLA); Gemini in Android Studio (UI mockups → Jetpack Compose).
- Jun 2025: Gemini 2.5 GA (Pro + Flash) + Flash-Lite; Gemini CLI open-source terminal agent.
- Nov–Dec 2025: Gemini 3 Pro announcement; 3 Deep Think in app; 3 Flash succeeds 2.5 Flash in positioning.
- Jan 2026: Apple × Google — Siri / Gemini partnership narratives in press.
- Feb–Mar 2026: Gemini 3.1 Pro, 3.1 Flash Image, 3.1 Flash Lite (per version table above).
Products & ecosystem
- Gemini Advanced — subscription tier (Google One AI Premium) bundling Ultra-class access.
- Gemini app — consumer surface; AI Studio & Vertex AI for builders.
- Jules — experimental GitHub-integrated coding agent (2.0 Flash Experimental era).
- Apple × Google (2026) — press coverage of Gemini-class models powering next-gen Siri experiences.
References
Primary encyclopedic entries: Gemini (language model) and Google Gemini. Related: Gemini 3 Pro, Gemma (language model), Google DeepMind. Technology pages: Gemini (DeepMind), Gemma models. Consumer hub: gemini.google.com.
All dates and statuses are compiled for this museum from Wikipedia and public Google posts as of March 2026; they are not an official Google product sheet.