SYSTEM Cited by 2 sources

Gemini

Gemini is Google DeepMind's multimodal LLM family, served as Google's primary consumer-facing AI assistant (gemini.google.com) and as the foundation-model substrate for Google product surfaces including Ask Maps, Ask YouTube, Antigravity, and AI Studio.

This is a minimum-viable wiki page anchored to the 2026-05-28 Google Research I/O roundup post. The post canonicalises Gemini along several axes:

• Operational scale: Gemini deployed in "more than 70 languages across more than 230 countries"; framed as "the most widely available AI assistant in the world" (Source: sources/2026-05-28-google-a-new-era-of-innovation-google-research-at-io-2026).
• Quality / factuality programme: Google Research has a multi-year factuality research arc (FACTS Grounding benchmark, text-to-image / video / long-context / uncertainty-expression factuality) feeding Gemini's factual-accuracy improvements (Source: sources/2026-05-28-google-a-new-era-of-innovation-google-research-at-io-2026).
• Latency programme: serving-infra extensions to speculative decoding — block verification + tree-structured drafting — power Gemini 3.5 Flash's current speed, implemented on TPU (Source: sources/2026-05-28-google-a-new-era-of-innovation-google-research-at-io-2026).
• Long-conversation capability: three named challenges — long-context reasoning, early-constraint adherence, longer RL trajectories — are claimed Google Research focus areas.

The page will deepen with future Google / Google DeepMind posts that decompose the Gemini family hierarchy (Pro / Flash / Nano tiers), training infrastructure, multimodal capabilities, and product-integration depth. Wiki-canonical Gemini variants:

• Gemini 3.5 Flash — the latency-optimised tier; serving target for current speculative-decoding extensions.

Seen in

• sources/2026-05-28-google-a-new-era-of-innovation-google-research-at-io-2026 — operational scale (70+ languages / 230+ countries); the serving-infra latency programme attribution to Gemini 3.5 Flash; multi-year factuality programme.
• sources/2026-06-02-instacart-semantic-ids-product-understanding-at-scale — Gemini Flash as the cheap LLM-attribute-extraction step in Instacart's discovery-flavor SID pipeline. Quote: "It first runs the product through Gemini Flash (~10x faster, ~5x cheaper than full-size models) to extract structured attributes (product type, key ingredients, dietary tags, format), stripping away marketing copy." Canonical wiki instance of Gemini Flash as a fast/cheap LLM preprocessing tier for representation-learning pipelines — the upstream of the Gemma-embedding step in the discovery flavor of Instacart Semantic IDs. (See patterns/llm-attribute-extraction-before-embedding.)

Related

• systems/gemini-3-5-flash — latency-optimised tier.
• systems/gemini-cli — Gemini command-line developer tool.
• systems/ai-studio — Gemini developer playground.
• systems/antigravity — Gemini developer environment.
• systems/ask-maps — Maps-integrated Gemini surface.
• systems/ask-youtube — YouTube-integrated Gemini surface.
• systems/google-tpu — Gemini's serving substrate.
• systems/gemma — sibling Google model family; co-deployed in Instacart's discovery-flavor SID pipeline.
• systems/instacart-semantic-ids — Gemini Flash as attribute-extraction upstream.
• concepts/speculative-decoding — Gemini-Flash inference primitive.
• concepts/factuality-decoding — sibling research arc impacting Gemini quality.
• concepts/precision-vs-discovery-codebook-flavor — Gemini Flash sits on the discovery side of this design axis.
• patterns/llm-attribute-extraction-before-embedding — pattern Gemini Flash exemplifies.
• patterns/two-flavor-codebook-precision-vs-discovery — the broader design pattern.
• companies/google — operator.
• companies/instacart — Gemini-Flash consumer.