Google officially declared war
2026YearGoogle I/OThe developer's conference, it's only two words to feel: "Crazy."。
It's not just..AISmart bodies, like ducks, are seamlessly plugged into all core traffic portals, such as search, browser, cell phone, smart glasses and so forthGemini 3.5 FlashVideo modelsOmniNewAIAssistantSparkThree kings blow。
After the musclesCut woodThe blogger adds:GeminiThe moon is broken9Millions; and synchronized with the announcement of substantial price reductions。
I can't put it straight:I'm better than you and cheaper than youI don't know。
What is this declaration of war
01
The most amazing thing in the General Assembly, no doubtGemini 3.5 FlashThe flash。
In general“..Pro"For the strengthFlash"for lightweight and fast。
In terms of model parameters3.5 FlashIt's actually less than3.1 PROBut in almost all the reasoning and coding benchmarking tests, the former perform better:
Complicated mathematical reasoningGSM8KTesting3.5 FlashGot it95.8%# Scores, over #3.1 PROYes93.2 per cent;code generation capabilitiesSWE-benchThis post is part of our special coverage Syria Protests 20113.5 FlashResolution rate achieved38.4%It's too much3.1 PROYes32.1%I'm sorry
Why
Based onDeepMindPublishedGemini 3.5 Technical ReportTwo of the most important core technologies。
Maximum knowledge distillationGoogle didn't train by simply countingFlashIt's the one that never came out“..Gemini 3.5 UltraIt's not the sameAs a teacher model, yesFlashDetachment。
Based onDeepMindChief ScientistJeff DeanIt's a tweet3.5 Flashfine-tuning of high-quality logical chain data setsI don't knowIt's better than the previous generation400%I don't know。
That means it inherited a supermodel“..Logical brainIt's not the sameIt's not a hard-on“..Knowledge baseIt's not the sameI don't know。
NewMoEStructure (mixed expert model)Other Organiser3.5 FlashInternally, Google uses a network of experts with a more finer particle size。
TraditionalMoEMaybe just..8or16A specialist, every time he activates1-2Each, a model that supports the size of hundreds of billions of parameters。
And based ona16zYes2026YearAIAnalysis of capital investment memorandums3.5 FlashYes256A mini-expert. Every reasoning can activate the most efficient of themFourOne。
That is why it is able to cover an extremely large multimodular feature space while maintaining very low activation parameters。
YesTTFTI'm sorryTime to First Token, first word output time) on the indicator3.5 FlashIt's done65Within milliseconds。
And humans needed it once100-150milliseconds。
In short, when it operates as an intelligent body, no pause can be detected from the biological perspective of human beings。
It is a truly perfect super-agent base for developers who require frequent access to tools, multiple rounds of reflection and very low delays。
Only with such extreme engineering optimization can it be established in an environment of intense competition“..Side downThe power of domination。
The first is the original modelGemini Omni FlashI don't know。
OmniIt means everythingGPT-4oYou can see how much gunpowder tastes。
At least in terms of performanceGemini Omni FlashFar aboveGPT-4oQualified fori'm sorry"This character。
EarlySoraorGemini 1.5It's basically a suture, which is the transliteration of words, the re-literation of text。
But..This timeOmniis the real root-to-end multi-modular alignment. Not only does it understand time consistency and physical patterns in video, but it also delays from400-600The industry average in milliseconds dropped to120milliseconds。
An example of a press conference: users pour water with cameras, water cups are running fullOmniBefore the spill0.5The second says, "Stop, stop!"
This real-time extrapolation of the physical state of the real world seems simple but significant:AIFormal evolution of chat robots from screen to reality support toolI don't know。
Although it is only a preliminary stage。
Second, smart assistantSparkI don't know。
Based onThe VergeInterviewsAndroidThe story of the vice president of engineeringSparkIt's been grantedAndroid 17Primary at the bottom of the systemAPIRight。
In short, you used to need a lot of lightAppIt's a complicated process to completeSparkOne, it can help you with everything, even sending information according to your tone, preferences, sorting out mail, aggregating schedules, tracking web dynamics, identifying hidden bill charges, batch processing documents, etc
In other words, laterAIAssistant, we don't need muchApp, any complex operation is simplified to be unique。
Third, smart glasses。
Why glasses again
At least in Google, seamless access to vision and hearing is the ultimate host of a large multi-modular model。
These glasses don't seem to have any fancy look, all focused on practical abilities:
Weight onlyFourShitMicro-OLEDFull-colour waveming lenses with a high radiance85%；
Carrying Light QuantitiesGeminiSide chip, local reasoning delayed12ms, real-time translation, image recognition, scene analysis without networking
Native connectionSparkSmart bodies, synchronized mobile phones, cloud data, personalized services such as calendar alerts, real-time translation, environmental alerts, etc。
In short, it's..Over the phone screenI don't knowPut your intelligence through your glasses into human first viewI don't know。
It's a lot of content, and Google seems to have cleaned up all the tricks and told the market the truth:
No entrance algorithm, nothing。
Parameters for scrolling large models, roll run pointsThe times have passed, the mere model providers have no moat, and the future is..End+Clouds+Ecology+Hardware's 4D Space War。
PutAIIn fact, it's reshaping the flow distribution logic of the Internet: From User Active SearchI'm sorryClick, becomes "AISmart body active distribution service”。
This could not have been better for a large number of developers and SMEs, as bottom computing and models had become extremely cheap, so that people could focus on innovation at the application level。
But the other competitors, I'm afraid, are trying to scold。
02
♪ When the wood is slashed on the stage ♪“..GeminiMonthly Active Users Official Breakout9Billions of dollars were created under the stage。
9Billions, more than any American rivalMAUMore than that。
How
The answer is simple and harsh:StuffI don't know。
Google doesn't need to be independentAIThe company just needs to pay for advertisingChomeAdd an icon next to the browser address bar in the30The bottom navigation column of the billion Andre mobile phones combines a pop-out shortcutGoogle WorkspaceSend updates in full..
The cost of getting a client is basically equal toCC BY-NC-ND 2.0I don't know。
And more importantly, for some time9The eyes and eyes of millions of active users looking at goods with smart glasses every daySparkAmended logic in dealing with matters andOmniThe interaction of visual models that produce high-quality, multimodular real-world feedback will all be nourishedGemini 4The nutrients。
This is an extremely strong barrier:The better the model works- WhatThe more people we use, the more people we use- WhatThe more data you get- WhatModels get betterI don't know。
In order to strengthen this ring quickly, Google has declared a price war directly to all its opponents:AI UltraPackage from249.99United States dollarsI'm sorryMonthCutPresent99.9United States dollarsI'm sorryMonthI don't know。
3.5 FlashMillionstokenThe input price is dry0.02Dollars, millionsTokenOutput price0.08United States dollars。
What kind of price is this
By comparison, the average prices of industry-level models are different0.15-0.2United States dollars and0.6-1United States dollars。
I'll take care of it every day. ApproximatelyOneTens of millionstokenI don't know. Put80%Task load cut toGemini 3.5 FlashLast year, it saved more than10Billion dollars。
How dare youAIFor cabbage
The biggest thing is:Computer infrastructure for vertical integrationI don't know。
IncludingOpenAII don't knowAnthropicThe giants, who look like the wind, are essentially “calculating renters” who need to buy power from Microsoft and Amazon, who are going to pay Wong。
And Google has its ownTPUPlus..3.5 FlashVery sickMoERare-activation efficiency reduces the calculation cost to a maximum。
It's all availableIt's an asset advantage to take down a simple algorithmI don't know。
The logic is clear。
The base large model is being rapidly commodified. Like water and electricity. What water company did you see
Google is not afraid that the large model itself will not make money because it can be earned by searching for advertising, cloud services and Andromeda ecology。
But for a big modelAPILifeOpenAII don't knowAnthropicI don't knowCohereI don't knowMistralThat's impossible。
The investors may now want to follow Ottoman's lead:“..GoogleAPIThe price is one tenth of your performance. Tell me, how does your business runIt's not the same
The pattern of competition in multiple industries will thus enter the accelerated shuffle period。
AIThere is no need for manufacturers to say that cheaper sources of computing must be found as soon as possible, or that they end up making chips themselves。
The next one is apples still behind closed doors。
Smart glasses+ OmniVideo Big Model+SparkPortfolio of primary system-level takeoversThere's no doubt about the threatiPhoneI don't know。
According to McGree's E-Trends in Consumption Report:For the next three years, based on visionI'm sorryShare of voice-free interaction, expected from current8%Jump to35%I don't know。
If the user gets used to using glasses and voiceDaily work and recreationTime used on screenAbsolutelyLarge compression。
If the apples don't bring in enough amazing clothing to fight backVision ProToo heavy and too expensive to be a few toys) whose access monopoly in the age of mobile Internet will be challenged unprecedentedly。
It's not an iterative, it's a revolution。
Google used three knives, with technology, traffic and prices, to put all its opponents in battle。
Is there anyone else laughing at it right now
