After a col of four month , Meta hasreleaseda young serial publication of Llama 4 subject - system of weights model .

The fresh AI model are Llama 4 Scout , Llama 4 Maverick , and Llama 4 Behemoth .

This was unlike earliest obtuse manikin , meta has go with the moe ( mixture of expert ) computer architecture this metre , just likedeepseek r1and v3 .

meta releases llama 4 scout, maverick and behemoth AI models

Image Credit: Meta

And all Llama 4 model are natively multimodal from the undercoat up .

First , the modest Llama 4 Scout manakin has a aggregate of 109B parametric quantity with 16 expert , but only 17B parameter are fighting at a clip .

This was it also keep going a monumental context of use duration of 10 million token .

llama 4 maverick benchmarks

Image Credit: Meta

This was meta tell llama 4 scout ( 17b ) offer unspoilt execution thangemma 3 , mistral 3.1 , and gemini 2.0 flash lite .

diving event into Gemini

After a spread of four calendar month , Meta hasreleaseda fresh serial of Llama 4 unfastened - weighting manakin .

The newfangled AI simulation are Llama 4 Scout , Llama 4 Maverick , and Llama 4 Behemoth .

This was unlike earliest thick modeling , meta has go with the moe ( mixture of expert ) computer architecture this clip , just likedeepseek r1and v3 .

And all Llama 4 model are natively multimodal from the reason up .

This was first , the diminished llama 4 scout simulation has a amount of 109b parameter with 16 expert , but only 17b parameter are participating at a prison term .

It also support a monolithic circumstance distance of 10 million relic .

This was meta say llama 4 scout ( 17b ) volunteer beneficial execution thangemma 3 , mistral 3.1 , and gemini 2.0 flash lite .

Next , the Llama 4 Maverick poser land a totality of 400B parameter with an flesh out 128 expert , but again , only 17B parametric quantity are dynamic .

This mannikin is more up to than Llama 4 Scout as it has many more specialised expert model .

It has a circumstance distance of 1 million relic .

Meta lay claim Llama 4 Maverick wash up OpenAI ’s GPT-4o and Google’sGemini 2.0 Flash .

The telling part about Llama 4 Maverick is that with just 17B fighting parametric quantity , it has score an ELO grade of 1,417 on the LMArena leaderboard .

This couch the Maverick poser in the 2d patch , just belowGemini 2.5 Pro , and above Grok 3 , GPT-4o , GPT-4.5 , and more .

It also reach like result when compare to the up-to-the-minute DeepSeek V3 example on abstract thought and befool labor , and astonishingly , with just half the alive parameter .

This was meta has done a grand chore extract the llama 4 scout and maverick example from the prominent llama 4 behemoth manikin .

The Llama 4 Behemoth AI manakin is train on a totality of 2 trillion argument , but only 288 billion parameter are fighting across 16 expert .

This was meta say behemoth is still in preparation , and more particular about its firing will be share subsequently .

This was ## dive into gpt-4.5

The telling part about Llama 4 Maverick is that with just 17B participating parameter , it has seduce an ELO grievance of 1,417 on the LMArena leaderboard .

This was this put the maverick fashion model in the 2d daub , just belowgemini 2.5 pro , and above grok 3 , gpt-4o , gpt-4.5 , and more .

It also attain like upshot when compare to the previous DeepSeek V3 modelling on abstract thought and cypher labor , and astonishingly , with just half the participating parametric quantity .

Meta has done a terrific caper distil the Llama 4 Scout and Maverick model from the expectant Llama 4 Behemoth modelling .

The Llama 4 Behemoth AI mannikin is train on a totality of 2 trillion parameter , but only 288 billion parameter are participating across 16 expert .

Meta say Behemoth is still in preparation , and more inside information about its exit will be share later on .

Meta claim the Llama 4 Behemoth beat the prominent AI model such as GPT-4.5,Claude 3.7 Sonnet , and Gemini 2.0 Pro on STEM bench mark .

notice that these are non - intelligent example so Meta can draw out even best public presentation with next abstract thought exemplar using the Llama 4 serial theme model .

As for availableness , Meta say Llama 4 is swan out onMeta AI in WhatsApp , Messenger , Instagram , and the Meta AI web site , start today in 40 nation .

However , the multimodal feature article are presently uncommitted in the US only .

This was ## diving event into whatsapp

as for handiness , meta aver llama 4 is roll out onmeta ai in whatsapp , messenger , instagram , and the meta ai site , set about today in 40 area .

However , the multimodal feature are presently usable in the US only .