Frontier llm-provider running a dual-track model strategy: the proprietary gemini line (API/product, closed weights) and the open-weight gemma-4 family (Apache-2.0, self-hostable). One of the few labs shipping at both ends — closed frontier and permissive open weights.
In this wiki
- gemini — the closed-weight frontier family. The May 2026 roundup google-ai-updates-may-2026 shipped Gemini 3.5 ("frontier intelligence for agents and coding"), Gemini Omni (multimodal video generation), and Gemini for Science.
- gemma-4 — the open-weight family, 2.3B to 31B gemma-4-technical-report. The 12B introduced encoder-free multimodality with native audio gemma-4-12b-announcement.
- gemma-variants — 16 domain-specialized models under the Gemma name (MedGemma, TxGemma, PaliGemma 2, VaultGemma, ShieldGemma 2 …), catalogued in awesome-gemma. The open track is not one lineup but a base Google keeps re-training per domain, across three live generations at once.
- Listed among the 2026 open-weight leaders in open-source-llms-2026.
Scale (Q2 2026, company-reported)
From alphabet-q2-2026-earnings: Google's model APIs process ~22B tokens/minute (up from ~16B a quarter earlier), the Gemini app has 950M MAU (DAU tripled Y/Y), 9M+ developers build on the APIs, and Gemma downloads total ~900M (Gemma 4 ~300M since April). This is the dual-track strategy showing up as usage volume on both the closed (gemini tokens) and open (gemma-4 downloads) tracks — the market expanding at the consumption layer, not only the capability layer.
The open track is a platform, not a product line
The dual-track framing below is about licensing. gemma-variants shows a second axis: on the closed track Google ships a handful of Gemini models and sells access, while on the open track it ships a base plus a widening set of domain forks it does not charge for at all — medicine, drug discovery, translation, privacy, interpretability, even animal communication. Reach is the return, and it is measurable: ~900M Gemma downloads alphabet-q2-2026-earnings.
Positioning
Unlike the open-only Chinese labs (deepseek, Qwen, Moonshot) or the closed-only frontier labs, Google straddles both: gemini holds the proprietary llm-benchmarks frontier while Gemma seeds the open-weight ecosystem (and the developer mindshare that comes with it). The May 2026 framing of both tracks around agents and coding shows the strategy converging on one use-case pitch across the open/closed divide.
Also an agent-platform vendor (cross-wiki)
Google's relevance extends past the model market into agent tooling: it ships the adk (Agent
Development Kit) and Gemini CLI, and is the second major vendor adopting the open agentskills-spec
after anthropic — the cross-vendor evidence that portable agent skills are an industry standard, not
one lab's house style. That thread is developed in ../agentic-tooling-wiki (this is the canonical Google
node both spokes link to).
Related
gemini · gemma-4 · llm-provider · open-weight-models · deepseek · anthropic · adk · agentskills-spec