--- title: Gemma 4 model family: sizes, architectures, and when to... description: Gemma 4 is a family of five open-weight, multimodal models (E2B, E4B, 12B, 26B/A4B, 31B) spanning four architectures, from efficient on-device designs using... canonical: https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Gemma 4 model family: sizes, architectures, and when to use open models | daily.dev og:description: Gemma 4 is a family of five open-weight, multimodal models (E2B, E4B, 12B, 26B/A4B, 31B) spanning four architectures, from efficient on-device designs using... og:url: https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft og:image: https://api.daily.dev/og/posts/36wtVfNft.png og:image:alt: Gemma 4 model family: sizes, architectures, and when to use open models og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Gemma 4 model family: sizes, architectures, and when to use open models **[Collections](https://daily.dev/sources/collections)** · 3 min read · 1 upvotes · 0 comments ## Summary Gemma 4 is a family of five open-weight, multimodal models (E2B, E4B, 12B, 26B/A4B, 31B) spanning four architectures, from efficient on-device designs using per-layer embeddings to a 26B mixture-of-experts model that activates only 4B parameters and a dense 31B flagship. The case for running these locally alongside frontier APIs like Gemini centers on cost, latency, and offline capability, with examples from Cue, HubX's BetterSpeak, K-Dense's Faraday, and Latitude showing gains from local inference on classification, summarization, and extraction tasks where raw reasoning power matters less than speed and cost. ## Content Gemma 4 is a family of multimodal, open models that comes in five sizes across four architectures. Which one you'd pick depends on where it has to run. The two smallest, E2B and E4B, are built for on-device use. They rely on per-layer embeddings (PLE) to stay efficient and ship with dedicated audio and image encoders. The 12B model takes a different route: it drops the dedicated encoders and projects audio straight into the LLM. The 26B model (A4B) is a Mixture-of-Experts design that activates only 4 billion parameters at a time, and it pairs that with a larger vision encoder for harder vision tasks. At the top, the 31B dense model is the most capable in the family and also uses the larger vision encoder. The models are released under Apache 2.0. ## Does a small model hold up in daily use? One person ran quantized 2B-4B parameter local models on an iPhone 16 for a month, and the result surprised me a little: roughly 90% of their everyday chatbot prompts, like rewording emails or explaining a term, never needed Claude or ChatGPT. They used Google AI Edge Gallery, which runs Gemma models in LiteRT format. It supports image and audio input and includes Agent Skills for tool use. The other 10% is worth taking seriously, though. Cloud models still won on research-heavy questions and on unfamiliar territory like coding. Extended use drained the battery quickly. And a local conversation can't follow you from one device to another, which is more of a nuisance than you'd expect. ## The same logic applies to startups Google Cloud makes the business version of this argument: pair frontier LLM APIs with compact open-weight models like Gemma 4, and you gain control over latency, infrastructure overhead, and the margin erosion that comes from sending every request to a paid API. Domain variants such as MedGemma exist for specialized work, and on-device and edge deployment are supported. The case studies they cite are Cue, HubX's BetterSpeak, K-Dense's Faraday, and Latitude. Between them they report lower latency, cost savings from running offline, and better retention after moving to what Google calls a compound AI stack. Put the two stories side by side and they say the same thing. Most requests are simple enough for a small model running close to the user. Keep the frontier model for the hard ones, and route the rest locally. ## Questions this post answers ### What are the different sizes and architectures in the Gemma 4 model family? Gemma 4 includes five models across four architectures: E2B and E4B are small on-device models using per-layer embeddings with dedicated audio and image encoders; 12B drops dedicated encoders and projects audio directly into the LLM; 26B (A4B) is a mixture-of-experts model with 26 billion total parameters but only 4 billion active at inference; and 31B is the largest, densest model, both 26B and 31B pairing with a larger vision encoder. _Comparing model architectures before picking one for a project is easier with daily.dev's developer community._ ### How many parameters does Gemma 4's 26B model actually activate during inference? The Gemma 4 26B model, also called A4B, only activates 4 billion of its 26 billion total parameters at inference time because it uses a mixture-of-experts architecture. This keeps compute costs manageable while still benefiting from a larger total parameter count, and it pairs with a larger vision encoder for harder visual tasks. _Developers weighing compute costs against capability can track model releases like this on daily.dev._ ### Why would a company run a smaller open model like Gemma locally instead of using a frontier API like Gemini for every request? Running smaller models locally cuts latency and cost for tasks that don't need frontier-level reasoning, such as classification, summarization, or structured extraction. Real examples include Cue reducing latency with local classification, HubX's BetterSpeak cutting costs and enabling offline use, K-Dense's Faraday avoiding per-request costs that would break its pricing model, and Latitude improving retention through lower latency. _Teams deciding between local inference and cloud APIs often compare tradeoffs like these on daily.dev._ ## Similar posts on daily.dev - [Google’s Gemma 4 shines on local systems – both big and small](https://daily.dev/posts/google-s-gemma-4-shines-on-local-systems-both-big-and-small-qpert58dk) · InfoWorld · 0 upvotes · 0 comments - [Welcome Gemma 4: Frontier multimodal intelligence on device](https://daily.dev/posts/welcome-gemma-4-frontier-multimodal-intelligence-on-device-mkjiapoth) · Hugging Face · 8 upvotes · 3 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#local-ai](https://daily.dev/tags/local-ai), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts), [#gemma](https://daily.dev/tags/gemma) [View this post on daily.dev](https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Gemma 4 model family: sizes, architectures, and when to use open models","url":"https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft"},"datePublished":"2026-09-28T16:02:37.178Z","dateModified":"2026-09-30T10:26:50.844Z","description":"Gemma 4 is a family of five open-weight, multimodal models (E2B, E4B, 12B, 26B/A4B, 31B) spanning four architectures, from efficient on-device designs using...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/068ffcf001fdb17bec225e96b81c4cf4?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/068ffcf001fdb17bec225e96b81c4cf4?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,local-ai,mixture-of-experts,gemma","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Gemma 4 model family: sizes, architectures, and when to use open models"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/gemma-4-model-family-sizes-architectures-and-when-to-use-open-models-36wtvfnft#faq","mainEntity":[{"@type":"Question","name":"What are the different sizes and architectures in the Gemma 4 model family?","acceptedAnswer":{"@type":"Answer","text":"Gemma 4 includes five models across four architectures: E2B and E4B are small on-device models using per-layer embeddings with dedicated audio and image encoders; 12B drops dedicated encoders and projects audio directly into the LLM; 26B (A4B) is a mixture-of-experts model with 26 billion total parameters but only 4 billion active at inference; and 31B is the largest, densest model, both 26B and 31B pairing with a larger vision encoder. Comparing model architectures before picking one for a project is easier with daily.dev's developer community."}},{"@type":"Question","name":"How many parameters does Gemma 4's 26B model actually activate during inference?","acceptedAnswer":{"@type":"Answer","text":"The Gemma 4 26B model, also called A4B, only activates 4 billion of its 26 billion total parameters at inference time because it uses a mixture-of-experts architecture. This keeps compute costs manageable while still benefiting from a larger total parameter count, and it pairs with a larger vision encoder for harder visual tasks. Developers weighing compute costs against capability can track model releases like this on daily.dev."}},{"@type":"Question","name":"Why would a company run a smaller open model like Gemma locally instead of using a frontier API like Gemini for every request?","acceptedAnswer":{"@type":"Answer","text":"Running smaller models locally cuts latency and cost for tasks that don't need frontier-level reasoning, such as classification, summarization, or structured extraction. Real examples include Cue reducing latency with local classification, HubX's BetterSpeak cutting costs and enabling offline use, K-Dense's Faraday avoiding per-request costs that would break its pricing model, and Latitude improving retention through lower latency. Teams deciding between local inference and cloud APIs often compare tradeoffs like these on daily.dev."}}]} ```