Google launches Gemma 4 12B, bringing multimodal AI agents to consumer laptops with 16GB RAM

Google's new Gemma 4 open AI model is sized for your laptop

The generative AI boom has driven the cost of memory into the stratosphere, and Google is a key part of that trend. So it's only fitting that Google should offer some less RAM-hungry local AI models. The company has announced the release of a new Gemma 4 model that fills a gap in the lineup that launched earlier this year. The new model is efficient enough that you may be able to run it on a pretty average consumer laptop. Back in April, Google released four models in the Gemma 4 family, which also marked the shift to a more open Apache 2.0 license. The initial models included two mobile-optimized options (E2B and E4B) along with a pair of models for more serious work (26B Mixture of Experts and 31B Dense). That left a rather large unserved space in the middle, which is right where the new model falls. Gemma 4 12B is considerably more capable than the mobile versions, but it won't require a $20,000 AI accelerator to run locally. Google says Gemma 4 12B is unique in that it can run on many consumer laptops without sacrificing quality. As long as you've got a computer with 16GB of system RAM or VRAM, the 12-billion-parameter model will work. That's about half the total memory footprint of Gemma 4 26B MoE, and Google claims the new model is almost as capable, at least as far as benchmarks go. Google says the new model is capable of complex multi-step reasoning and agentic workflows that previously required the larger Gemma variants. Despite the smaller parameter count, Gemma 4 12B comes with the newly devised Multi-Token Prediction (MTP) drafters, which take advantage of unused processing cycles to calculate possible future tokens. The result is greater speed and efficiency. Google has released optional MTP versions of the other Gemma 4 models, but this is the first one to have MTP out of the box. Gemma 4 12B is also more efficient thanks to a new approach to multimodality. The Gemma 4 family is natively multimodal, accepting text, audio, or images as inputs. Most gen AI models -- including the other Gemma 4 variants -- use dedicated encoders to process non-text inputs and pass that data to the LLM. This works well enough, but it increases latency and memory usage. With the new mid-weight model, Google has implemented a streamlined embedding module for vision, featuring single-matrix multiplication and positional embedding, which allows the data to pass to the LLM with proper spatial awareness. This eliminates the need for a bulky middleman encoder. For audio, there's no encoding at all. The developers worked out a method of projecting the raw audio signal into the same vectors used for text tokens. If you want to check out the new Gemma 4 model, it's accessible without a download via tools like LM Studio Google AI Edge Gallery, and more. But the whole idea with Gemma 4 12B is that you can run it locally and on your own terms. If you've got the RAM, the model weights are available for download immediately on Kaggle and Hugging Face. It's just shy of 18 GB.

Google brings local AI agents to laptops with Gemma 4 12B

In a blog post, the company said the model, combined with the Google AI Edge stack, can be used to build and test applications on everyday machines. The model-runtime combination supports capabilities such as autonomous data processing, visual insight generation, webpage creation, and tool use. The release includes Google AI Edge Gallery for macOS, where developers can use Gemma 4 12B to generate and run scripts for tasks such as data analysis. Google also said its Eloquent voice dictation and editing app now runs fully on-device on macOS, with support for local transcription and voice-driven text editing. Google has also expanded LiteRT-LM, its lightweight command-line tool for running language models locally, with a new serve command. The company said this allows the CLI to act as a local LLM server and lets developers connect Gemma 4 12B to standard tools, SDKs, and frameworks through a local endpoint. "Your data stays on your device while maintaining reliable responsiveness, utility, and cost efficiency," the company said in the blog post. The announcement comes as enterprises are looking beyond large, general-purpose models for some AI workloads. Gartner predicted that by 2027, organizations will use small, task-specific AI models at least three times more than general-purpose large language models, citing demand for more contextualized and cost-effective AI systems.

Google's latest on-device AI model is custom-made for your laptop

It utilizes an encoder-free architecture to offer multimodal performance without the latency introduced by encoders. The new model performs close to the Gemma 4 26B MoE model in benchmarks. Back in April, Google released its mobile-friendly Gemma E2B and E4B models, bringing on-device multimodal AI to Android and iOS devices. It also released the high-end 26B Mixture of Experts (MoE) and 31B Dense models for higher-end devices with dedicated AI GPUs. Now, the company is launching another Gemma model that sits nicely between the four. Google today announced the Gemma 4 12B model aimed at bringing on-device AI capabilities to laptops. It offers multimodal features and is the first mid-sized model from Google to support native audio input. The company claims that its 12B model delivers performance similar to the 26B MoE model in benchmarks, while being small enough to run on normal consumer laptops with 16GB of RAM. To achieve this, the company came up with unique solutions for supporting multimodal inputs without increasing latency and memory usage. Gemma 4 12B uses an encoder-free architecture to avoid the memory costs associated with encoders that are typically used in most multimodal AI models. For vision, it's using a lightweight module that utilizes "single matrix multiplication, positional embedding, and normalizations," allowing image data to be passed to the LLM without requiring an encoder in the middle. It also completely does away with encoding for audio inputs. Google was able to project the raw audio signal directly into the same dimensional space as text tokens. What that means is that Gemma 4 12B can handle multimodal inputs, just like the other Gemma models, but without the added overhead of encoding such inputs. This should result in much better performance on laptops without the need for dedicated AI hardware. Interested users can try the new model right now in LM Studio, Ollama, Google AI Edge Gallery, and more. If you're interested in running it locally on your laptop, the weights are available to download from Hugging Face and Kaggle.

Google AI Edge Gallery launches to macOS

In addition to Google AI Edge Gallery, which lets users run Gemma models locally on their Macs, the company also released the Gemma 4 12B model and the Google AI Edge Eloquent dictation app for the Mac. Here are the details. A bit of background The majority of users who rely on LLMs for everyday tasks tend to use ChatGPT, Claude, or Gemini, which are cloud-based models running on OpenAI, Anthropic, and Google's servers. Another way to interact with LLMs is through local models. These are usually much smaller and less capable than the trillion-parameter models that run in the cloud, but they also come with several advantages. For one, being less capable than cloud-based models does not mean they are bad. Also, they do not require an active internet connection, since they run on the computer's own processing power. Additionally, the better the computer, the faster the responses, and the larger the models it can handle. And finally, because everything runs locally, these models are more private too, since conversation data does not need to leave the device. There are a few ways to install local models on a Mac, and we covered this here, when OpenAI released its own open models. But in a nutshell, you need to install platforms such as Ollama and LM Studio, and then install a model that can runs smoothly on your Mac's hardware. Hugging Face hosts thousands of open models to choose from, including those from frontier labs. However, platforms such as Ollama and LM Studio also offer ways to install these models directly from them. Which brings us to Google AI Edge Gallery, Google's platform for running AI models locally. Google already offered a Google AI Edge Gallery app for Android and for iOS, but today the company released it for macOS as well. Google AI Edge Gallery and Gemma 4 12B One thing to note right from the get-go is that, contrary to Ollama and LM Studio, which allow users to install any AI model compatible with their hardware, Google AI Edge Gallery for Mac currently only offers access to 5 of Google's own models, where 'it' stands for instruct, meaning they can be tuned to follow user instructions rather than simply complete text: * Gemma-4-12B-it * Gemma-4-E2B-it * Gemma-4-E4B-it * Gemma-3n-E2B-it * Gemma-3n-E4B-it The top item on the list is particularly notable. Gemma 4 12B was released today, and it was designed to bring agentic, multimodal intelligence directly to your laptop," according to Google. While most consumer-facing local models from frontier AI labs tend to stay somewhere between 2 billion and 9 billion parameters, Google says Gemma 4's 12-billion-parameter design delivers performance comparable to its 26-billion-parameter mixture-of-experts model, while still being "small enough to run locally on consumer laptops with 16GB of RAM." Gemma 4 12B is also multimodal, which means it can handle text, vision, and audio. Google says that the model also packs good coding capabilities, "allowing you to extract meaningful insights from your data right on your device." You can learn more about Google AI Edge Gallery here, and you can learn more about Gemma 4 12B here. Google AI Edge Eloquent Alongside Gemma 12B and the release of Google AI Edge Gallery for macOS, Google also launched the Google AI Edge Eloquent app for Mac today, after bringing the app to iOS a few months ago. Google AI Edge Eloquent is a free dictation app that captures what users say and transcribes it while polishing the text, removing disfluencies, and making light edits for clarity and flow. Processing is done on-device, rather than on the cloud. The app also lets users choose between different writing styles and add custom words, such as names, jargon, and other terms they use often. That helps avoid the kind of frequent miscorrections that dictation apps can otherwise make with specific words and phrases. You can learn more about Google AI Edge eloquent here. Worth checking out on Amazon

Google's new open source Gemma 4 12B analyzes audio, video -- and runs entirely locally on a typical 16GB enterprise laptop

While many AI open source model providers are pursuing larger and more powerful models, Google is still giving attention to the smaller, more local side of the market. Today, the tech giant released Gemma 4 12B, an 11.95-billion-parameter open-weights model with permissive Apache 2.0 license optimized to execute locally on a standard enterprise laptop using just 16GB of VRAM or unified memory. That means those enterprise users looking to keep working with AI while on a flight without WiFi, or trying to keep it offline for security reasons, can now do so far more easily and at far less cost (free to download and operate). Gemma 4 12B's most notable breakthrough is an encoder-free "Unified" architecture, which allows raw audio waveforms and visual patches to flow directly into the core LLM backbone without the latency or memory overhead of secondary processing modules. Available immediately for download on Hugging Face and Kaggle and for use on Google AI Edge Gallery, Gemma 4 12B packs a 256K token context window, native agentic tool-use capabilities, and an explicit step-by-step reasoning mode into a highly optimized footprint that bridges the gap between mobile edge models and heavy data-center infrastructure. The Architectural Shift: Understanding the Encoder-Free Advantage Gemma 4 12B is highly relevant to enterprise architecture due to its novel "Unified" structure. Traditional multimodal systems typically utilize discrete, separate encoders to translate audio waveforms and visual data into representations that the core language model can process. This conventional approach inherently increases both inference latency and total memory consumption. Gemma 4 12B radically alters this pipeline by functioning entirely without these secondary encoders. Instead, visual patches and raw audio waveforms are projected directly into the core large language model's embedding space through lightweight linear layers. The vision encoder is replaced by a 35-million-parameter module utilizing a single matrix multiplication, while the audio encoder is eliminated entirely. For enterprise engineering teams, this unified architecture delivers distinct operational advantages: lower latency for multimodal tasks, reduced VRAM requirements (down to 16GB -- typical for laptops), and the ability to fine-tune the entire multimodal system in a single, cohesive pass. Performance Metrics and Core Capabilities Despite its compact size, Gemma 4 12B achieves benchmarks nearing Google's larger 26B Mixture-of-Experts model. Beyond static benchmarks, the model supports a massive 256K token context window. This is critical for enterprises needing to process lengthy financial reports, extensive code repositories, or hour-long meeting transcripts. Furthermore, Gemma 4 12B includes a native "thinking" mode to map out step-by-step reasoning before generating a response. It also features out-of-the-box support for native function calling and system prompts, which are essential prerequisites for building highly capable autonomous software agents. The Enterprise Verdict: Should You Adopt Gemma 4 12B? The short answer is yes, provided your operational needs align with edge computing, strict data privacy, or agentic automation. However, adoption should not be a blanket replacement for all existing AI infrastructure. Instead, technical leaders should view Gemma 4 12B as a specialized tool optimized for specific deployment conditions. * Strict Data Privacy and Compliance Mandates: Many enterprises operate in highly regulated sectors -- such as healthcare, finance, or defense -- where transmitting sensitive data, proprietary code, or confidential internal documents to third-party APIs is unacceptable. Because Gemma 4 12B is small enough to run locally on machines equipped with just 16GB of VRAM or unified memory, organizations can process sensitive multimodal data entirely on-premises or directly on employee laptops. This local execution eliminates the risk of data leakage and ensures compliance with strict regulatory frameworks. * Multimodal Autonomous Agent Workflows: If your engineering roadmap involves autonomous agents interacting with real-world inputs, Gemma 4 12B is uniquely positioned to serve as the reasoning engine. The combination of native function calling, robust coding capabilities, and the capacity to ingest real-time audio and variable-resolution images makes it highly suitable for agentic tasks. Google has simultaneously released a dedicated Gemma Skills Repository to explicitly support agentic development with these new models. * Cost-Sensitive Edge Deployments: For applications operating at the edge -- such as retail inventory monitoring via cameras, localized customer service kiosks, or offline field-service applications -- maintaining a persistent cloud connection is costly and sometimes impossible. The encoder-free architecture significantly lowers the total cost of ownership by reducing the hardware threshold needed for inference. Deploying a highly capable 12B model locally avoids recurring API costs and unpredictable cloud compute billing. When to Consider Alternative Solutions While Gemma 4 12B is powerful, it has specific constraints that technical leaders must acknowledge. * Massive Knowledge Retrieval: Like all large language models, Gemma 4 12B is a reasoning engine, not a static database. If your primary use case relies on vast, generalized factual retrieval without leveraging a robust Retrieval-Augmented Generation pipeline, you may still require larger foundation models. * Extended Video and Audio Processing: The model has hard limits on media ingestion. Audio inputs are strictly capped at 30 seconds of processing, and video understanding is limited to 60 seconds (assuming a processing rate of one frame per second). Enterprises looking to process feature-length videos or massive audio archives natively will hit bottlenecks and should consider API-based models or chunking architectures. Implementation and Ecosystem Readiness One of the strongest arguments for enterprise adoption is the model's immediate compatibility with the broader open-source development ecosystem. Google has ensured that Gemma 4 12B is not an isolated experiment; it is ready for production. Weights are available on Hugging Face and Kaggle, and the model integrates seamlessly with industry-standard deployment frameworks such as vLLM, SGLang, MLX, and llama.cpp. For organizations deeply embedded in Google Cloud, endpoints can be spun up quickly using the Gemini Enterprise Agent Platform Model Garden, Cloud Run, or Google Kubernetes Engine. For enterprise leaders aiming to decentralize their AI workloads, Gemma 4 12B offers a rare combination of edge-friendly efficiency and frontier-class reasoning. If your organization requires highly private, multimodal processing without the latency and cost of cloud reliance, Gemma 4 12B should be heavily evaluated for your next production pipeline.

See what 3 builders are making with Gemma 4

We recently released Gemma 4, our most capable open models to date. Since then, they have been downloaded more than 150 million times, and we've been expanding the family's capabilities. We introduced Multi-Token Prediction (MTP) to accelerate inference, and recently released the 12B Unified model and Quantization-Aware-Training (QAT) checkpoints. Released under an Apache 2.0 license, Gemma 4 gives builders and organizations flexibility to fine-tune and deploy models across a variety of environments, from edge devices to local workstations. Many builders are sharing what they've created with Gemma 4, showcasing how the models' capabilities translate into real-world applications. Here are three highlights of what people and companies are creating. Build low-latency, on-device apps. The team at the app building company HubX used Gemma 4 to build BetterSpeak, an offline AI English tutoring platform. BetterSpeak uses the edge-optimized Gemma 4 E2B (effective 2B parameters) model as the reasoning engine for its on-device pipeline, enabling private, low-latency tutoring without the need for an internet connection. To overcome mobile hardware constraints, HubX deployed the 4-bit quantized version of the model released by Google. This version handles tasks like grammar explanations and progress monitoring across multiple languages. By leveraging Gemma 4's native audio input capabilities, the app supports direct speech-to-speech learning, reducing costs while ensuring user privacy by processing all vocal and text data entirely on-device.

Introducing Gemma 4 12B: a unified, encoder-free multimodal model

Today, we are introducing Gemma 4 12B, our latest model designed to bring agentic multimodal intelligence directly to laptops. Bridging the gap between our edge-friendly E4B and our more advanced 26B Mixture of Experts (MoE), Gemma 4 12B packages powerful capabilities inside a reduced memory footprint. It is also our first mid-sized model to feature native audio inputs. Thanks to the developer community, Gemma 4 models have now crossed 150 million downloads. You've built everything from wearable robotic arms for physical assistance to enterprise-grade AI security. We're excited to see what you build with this latest addition. Here's an overview of what makes Gemma 4 12B unique: * Novel unified architecture: No multimodal encoders. The vision and audio inputs flow directly into the LLM backbone. * Advanced reasoning: Benchmark performance nearing our 26B model, unlocking powerful multi-step reasoning and agentic workflows. * Laptop ready: Small enough to run locally with just 16GB of VRAM or unified memory. * Open and accessible: Released under an Apache 2.0 license with support across the developer ecosystem. * Drafter-ready: Gemma 4 12B comes equipped with Multi-Token Prediction (MTP) drafters to reduce latency. Together, these features bring advanced multimodal capabilities to everyday hardware without sacrificing speed or reasoning. Let's now take a closer look at how Gemma 4 12B achieves this. Run state-of-the-art agents locally Gemma 4 12B delivers performance nearing our larger 26B MoE model on standard benchmarks, but at less than half the total memory footprint. Small enough to run locally on consumer laptops with 16GB of RAM, it unlocks powerful multimodal and agentic experiences right on your machine.

Google unveils Gemma 4 12B for local AI agents, coding, and multimodal reasoning

Google DeepMind has introduced Gemma 4 12B, a new open-weight multimodal model designed to bring agentic intelligence directly to laptops with mobile-first efficiency and advanced reasoning. Gemma 4 12B sits between the edge-friendly E4B model and the larger 26B Mixture-of-Experts (MoE) model, offering strong performance with a reduced memory footprint. It is also the first mid-sized model in the series to feature native audio input support. The Gemma family has now crossed 150 million downloads, with developers building use cases ranging from wearable robotic arms for physical assistance to enterprise-grade AI security systems. Key features of Gemma 4 12B Gemma 4 12B introduces a unified encoder-free multimodal architecture, where vision and audio inputs flow directly into the LLM backbone without separate encoders. This reduces latency and memory overhead compared to traditional multimodal systems. * Vision processing: Replaces the vision encoder with a lightweight embedding module using a single matrix multiplication, positional embeddings, and normalizations * Audio processing: Removes the audio encoder entirely and projects raw audio signals into the same token space as text The model delivers benchmark performance close to the 26B MoE model while using less than half the memory footprint, enabling multi-step reasoning and agentic workflows on laptops with 16GB VRAM or unified memory. Gemma 4 12B is released under the Apache 2.0 license and includes Multi-Token Prediction (MTP) drafters to improve inference speed and reduce latency. It supports advanced agentic capabilities such as: * Autonomous data processing * Generating rich visual insights * Building fully functional webpages * Executing everyday tool use and workflows * Multi-step reasoning and structured task execution A new Gemma Skills Repository is also introduced, providing an official library of reusable skills designed specifically for building agentic systems with Gemma models. Run state-of-the-art agents locally Gemma 4 12B delivers near-26B MoE performance on benchmarks while requiring significantly lower memory, making it suitable for: * Local AI agents * On-device reasoning systems * Private offline workflows * Edge and laptop-based AI applications Experience a uniquely efficient unified architecture Traditional multimodal systems rely on separate encoders for vision and audio, which increases latency and memory usage. Gemma 4 12B removes this limitation through a fully unified design. * No separate encoders for vision or audio * Direct processing inside LLM backbone * Reduced latency and memory consumption * Improved cross-modal reasoning consistency Vision pipeline Vision is handled through a lightweight embedding module with a single matrix multiplication, positional embeddings, and normalizations, replacing the full vision encoder. Audio pipeline Audio is processed by removing the encoder entirely and projecting raw audio signals directly into the same embedding space as text tokens. Performance Benchmarks Gemma 4 12B shows performance differences across Linux and macOS GPU environments, measuring prefill speed, decode speed, latency, and memory usage. Linux * Device: AMD Radeon™ AI PRO R9700 * Backend: GPU * Prefill: 662.32 tokens/sec * Decode: 66.26 tokens/sec * Time-to-first-token: 1.56 sec * Model size: 6235 MB * GPU Memory: 8064.2 MB macOS * Device: MacBook Pro M4 * Backend: GPU * Prefill: 243.55 tokens/sec * Decode: 29.56 tokens/sec * Time-to-first-token: 4.2 sec * Model size: 6235 MB * GPU Memory: 7763 MB Get started today Developers can try Gemma 4 12B using: * LM Studio * Ollama * Google AI Edge Gallery * Google AI Edge Eloquent * LiteRT-LM They can also: * Download weights from Hugging Face and Kaggle * Review developer documentation and quick start notebook * Use frameworks like Hugging Face Transformers, llama.cpp, MLX, SGLang, and vLLM * Fine-tune using Unsloth * Spin up production endpoints using Google Cloud Gemma Skills Repository The model includes an official Skills Repository, designed to help developers build agentic systems using reusable Gemma capabilities. Bringing Gemma 4 12B to your laptop Gemma 4 12B is designed for local execution on everyday machines using the Google AI Edge stack. This enables: * Autonomous data processing * Generating rich visual insights * Building fully functional webpages * Everyday tool execution * Fully local agent workflows Coding and advanced workflows Gemma 4 12B supports advanced local execution capabilities including: * Python code generation from natural language prompts * Local execution of scripts and data analysis * Automatic chart generation from datasets * Self-correcting code generation in a single turn * Complex 3D rendering tasks with dependency handling * End-to-end webpage generation In coding tests, the model can generate outputs such as charts from datasets (e.g., comparing baby names across years) and even render 3D scenes with full dependency setup and correction in a single prompt. Dictation and voice-driven editing Google AI Edge Eloquent is a fully on-device macOS application that transforms speech into structured writing. It provides: * System-wide voice dictation via hotkeys * Fully local transcription of audio and video files * Voice-based text editing (Voice Edit feature) Users can issue commands such as: * "Restructure these notes into an executive summary" * "Translate this into Hindi" Gemma 4 12B improves instruction following, scope adherence, and output quality compared to previous models, with a reported 60%+ improvement. LiteRT-LM and local serving LiteRT-LM introduces a new serve command, turning it into a drop-in local LLM server. This allows: * Standard API endpoints for local models * Drop-in replacement for hosted LLM servers * Integration with tools like Continue, Aider, OpenCode, Hermes, and Pi * Fully local agent workflows * Zero-code model deployment Deployment options Gemma 4 12B can be deployed across: * LM Studio, Ollama, Hugging Face Transformers, llama.cpp, MLX, SGLang, vLLM * Google Cloud endpoints * Gemini Enterprise Agent Platform Model Garden * Cloud Run and Google Kubernetes Engine (GKE) Availability Gemma 4 12B is available as an open-weight model under the Apache 2.0 license and can be downloaded from Hugging Face and Kaggle. It is optimized for laptops with 16GB memory and supports fully offline multimodal AI workflows. It is integrated across the Google AI Edge ecosystem, including macOS tools such as AI Edge Gallery, Eloquent, and LiteRT-LM CLI, enabling local-first AI experiences while keeping all data on-device.

Google unveils Gemma 4 12B, a local AI model for everyday PCs: Here is what it can do

It is designed to bring agentic multimodal intelligence directly to laptops. Google has introduced a new artificial intelligence model called Gemma 4 12B. The tech giant describes Gemma 4 12B as a "unified transformer" which is designed to bring agentic multimodal intelligence directly to laptops. The new model sits between the smaller Gemma E4B and the advanced 26B Mixture of Experts (MoE) model, offering a balance of performance and efficiency. Google also revealed that the Gemma family of models has crossed 150 million downloads. The company said developers have already used Gemma models for a wide range of projects, from wearable robotic arms for physical assistance to enterprise-grade security solutions. Google Gemma 4 12B: Key capabilities One of the biggest highlights of Gemma 4 12B is that it can run locally on devices with just 16GB of RAM or VRAM. According to Google, the model delivers advanced reasoning abilities while maintaining a relatively small memory footprint. Google also says Gemma 4 12B is its first mid-sized model with native audio input support. Also read: Microsoft AI chief says future artificial intelligence should help humans, not replace them Unlike many multimodal AI models that depend on separate encoders for visual and audio information, Gemma 4 12B handles these inputs directly through its language model backbone. Google describes this as a more streamlined approach that helps reduce memory usage and improve response speed. For image processing, Google replaced the traditional vision encoder with a lightweight embedding module. "This allows the LLM backbone to take over visual processing," Google explained. Also, instead of using a dedicated audio encoder, Gemma 4 12B projects raw audio signals directly into the same space used for text tokens. Google also highlighted that Gemma 4 12B comes equipped with Multi-Token Prediction (MTP) drafters to reduce latency. According to the company, the model delivers benchmark performance close to its larger 26B counterpart. This could make advanced multimodal AI and agent-based workflows more accessible to developers and users who want to run AI locally on everyday hardware.