HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Genie world modelsGoogle DeepMind—VeoGoogle DeepMind—GeminiGoogle DeepMind—Tensor Processing Unit (TPU)Google DeepMind—AlphaFoldGoogle DeepMind—Genie world modelsGoogle DeepMind—VeoGoogle DeepMind—GeminiGoogle DeepMind—Tensor Processing Unit (TPU)Google DeepMind—GeminiGoogle DeepMind—

Gemini

by Google DeepMindGBUS

ScalingStage 5 of 5

Gemini 3.6 Flash shipped in July 2026. Gemini 3.5 Pro, announced in May, is still delayed.

Updated 21 Jul 2026Checked 25 Sep0 updates this week

Milestones

Next · Gemini 3.5 Pro general release
  1. Gemini 1.0 announced6 Dec 2023Complete.
  2. Gemini 1.5 with 1M-token context15 Feb 2024Complete.
  3. Gemini 2.0 Flash11 Dec 2024Complete.
  4. Gemini 2.5 Pro reasoning model25 Mar 2025Complete.
  5. Gemini Deep Think reaches IMO gold-medal standard21 Jul 2025Complete.
  6. Gemini 3 Pro released18 Nov 2025Complete.
  7. Gemini 3.6 Flash and 3.5 Flash-Lite released21 Jul 2026Complete.
  8. Gemini 3.5 Pro general releaseNowCurrent milestone.

Most important updates

  • 21 Jul 2026
  • 19 May 2026
  • 18 Nov 2025
  • 21 Jul 2025
  • 6 Dec 2023

Upcoming

  1. Q4 2026Gemini 3.5 Pro release (next)

Current obstacles

  • Cost at Google scaleGemini runs inside Search for billions of users, so every upgrade is weighed against running cost and speed.
  • Wrong search answersAI Overviews have made high-profile mistakes. Reliably basing answers on real sources is still hard.

Physics limits

  • Attention cost grows with length squaredIn standard attention every token is compared with every other, so doubling the input quadruples that work. Very long contexts need shortcuts that can miss details.
  • All experts must sit in memoryMixture-of-experts saves arithmetic, not memory: a 1-trillion-parameter model needs about 1 TB at 8-bit precision even if only 3% runs per word, so it needs many linked chips.
  • Returns shrink as a power lawError falls only as a small power of training compute: each fixed step of improvement needs roughly 10x more compute, energy and money than the one before.

How it works

4 parts
Google's data centre campus in Mayes County, Oklahoma: Gemini is trained and served on TPU pods in buildings like these
Google's data centre campus in Mayes County, Oklahoma: Gemini is trained and served on TPU pods in buildings like thesePhoto: Xpda · CC BY-SA 4.0 (opens commons.wikimedia.org)
Inputs

One model, all media

Images, audio and video are chopped into the same kind of pieces as text, so one model can answer questions about an hour of video.

Experts

Mixture of experts

The model is split into many 'expert' sub-networks and a router activates only a few per word, so total size grows without the cost per word growing.

Context

Very long inputs

Gemini can read a million or more tokens (word pieces) at once, such as a whole codebase or a book.

Thinking

Deep Think

For hard maths and code, it tries several lines of reasoning in parallel and combines them.

Update log

5 updates

Tue 21 Jul

  • Minor: BlogSoftware

Tue 19 May

  • Minor: BlogSoftware

Tue 18 Nov 2025

  • Major: BlogSoftware

Mon 21 Jul 2025

  • Major: BlogResearch

Wed 6 Dec 2023

  • Major: BlogSoftware

About Google DeepMind

The team behind Gemini

Google DeepMind

Gemini, AlphaFold, Genie, Veo, TPU chips

Google DeepMind is Alphabet's AI lab, formed in 2023 from DeepMind and Google Brain. It builds the Gemini models and leads in AI for science. AlphaFold won a share of the 2024 Nobel Prize in Chemistry.

  • Gemini 3 (Nov 2025) led many benchmarks. Gemini 3.5 Pro, announced in May 2026, is delayed.
  • AlphaFold predicted structures for over 200 million proteins and won a share of the 2024 Chemistry Nobel.
  • Google's own Ironwood TPU chip is used in-house and sold to outside labs, including Anthropic.

deepmind.google (opens deepmind.google)Company profile →

Founded
201016 yrs
Headquarters
United KingdomUnited States
Status
Subsidiaryof Alphabet
Valuation
Alphabet-owned
Works in
AIFrontier modelsComputingBiotech
Coverage
5 programs · 30 updateslatest 29 Jul 2026checked 25 Sep
Customers
Anthropic
Partners
Isomorphic LabsBroadcom
Acquired by
Alphabet
People
Demis HassabisCEO and co-founderShane LeggCo-founder and Chief AGI ScientistJohn JumperAlphaFold lead; 2024 Nobel laureate
Primer