Google Unveils Gemma 3: A Powerful, Efficient AI Model for Single-GPU Applications

Google's Gemma 3 is an open source, single-GPU AI with a 128K context window

Most new AI models go big -- more parameters, more tokens, more everything. Google's newest AI model has some big numbers, but it's also tuned for efficiency. Google says the Gemma 3 open source model is the best in the world for running on a single GPU or AI accelerator. The latest Gemma model is aimed primarily at developers who need to create AI to run in various environments, be it a data center or a smartphone. And you can tinker with Gemma 3 right now. Google claims Gemma 3 will be able to tackle more challenging tasks compared to the older open source Google models. The context window, a measure of how much data you can input, has been expanded to 128,000 from 8,192 tokens in previous Gemma models. Gemma 3, which is based on the proprietary Gemini 2.0 foundation, is also a multimodal model capable of processing text, high-resolution images, and even video. Google also has a new solution for image safety called ShieldGemma 2, which can be integrated with Gemma to help block unwanted images in three content categories: dangerous, sexual, or violent. Most of the popular AI models you've heard of run on collections of servers in a data center, filled to the brim with AI computing power. Many of them are far too large to run on the kind of hardware you have at home or in the office. The release of the first Gemma models last year gave developers and enthusiasts another low-hardware option to compete with the likes of Meta Llama3. There has been a drive for efficiency in AI lately, with models like DeepSeek R1 gaining traction on the basis of lower computing costs.

Google calls Gemma 3 the most powerful AI model you can run on one GPU

Richard Lawler is a senior editor following news across tech, culture, policy, and entertainment. He joined The Verge in 2021 after several years covering news at Engadget. A little over a year after releasing two "open" Gemma AI models built from the same technology behind its Gemini AI, Google is updating the family with Gemma 3. According to the blog post, these models are intended for use by developers creating AI applications capable of running wherever they're needed, on anything from a phone to a workstation with support for over 35 languages, as well as the ability to analyze text, images, and short videos.

Google claims Gemma 3 reaches 98% of DeepSeek's accuracy - using only one GPU

The economics of artificial intelligence have been a hot topic of late, with startup DeepSeek AI claiming eye-opening economies of scale in deploying GPU chips. Two can play that game. On Wednesday, Google announced its latest open-source large language model, Gemma 3, came close to achieving the accuracy of DeepSeek's R1 with a fraction of the estimated computing power. Using "Elo" scores, a common measurement system used to rank chess and athletes, Google claims Gemma 3 comes within 98% of the score of DeepSeek's R1, 1338 versus 1363 for R1. Also: First Gemini, now Gemma: Google's new, open AI models target developers That means R1 is superior to Gemma 3. However, based on Google's estimate, the search giant claims that it would take 32 of Nvidia's mainstream "H100" GPU chips to achieve R1's score, whereas Gemma 3 uses only one H100 GPU. Google's balance of compute and Elo score is a "sweet spot," the company claims. In a blog post, Google bills the new program as "the most capable model you can run on a single GPU or TPU," referring to the company's custom AI chip, the "tensor processing unit." "Gemma 3 delivers state-of-the-art performance for its size, outperforming Llama-405B, DeepSeek-V3, and o3-mini in preliminary human preference evaluations on LMArena's leaderboard," the blog post relates, referring to the Elo scores. "This helps you to create engaging user experiences that can fit on a single GPU or TPU host." Google's model also tops Meta's Llama 3's Elo score, which it estimates would require 16 GPUs. (Note that the numbers of H100 chips used by the competition are Google's estimate; DeepSeek AI has only disclosed an example of using 1,814 of Nvidia's less-powerful H800 GPUs to server answers with R1.) More detailed information is provided in a developer blog post on HuggingFace, where the Gemma 3 repository is offered. The Gemma 3 models, intended for on-device usage rather than data centers, have a vastly smaller number of parameters, or neural "weights," than R1 and other open-source models. Generally speaking, the greater the number of parameters, the more computing power is required. Also: What is DeepSeek AI? Is it safe? Here's everything you need to know The Gemma code offers parameter counts of 1 billion, 4 billion, 12 billion, and 27 billion, quite small by today's standards. In contrast, R1 has a parameter count of 671 billion, of which it can selectively use 37 billion by ignoring or turning off parts of the network. The main enhancement to make such efficiency possible is a widely used AI technique called distillation, whereby trained model weights from a larger model are extracted from that model and inserted into a smaller model, such as Gemma 3, to give it enhanced powers. The distilled model is also run through three different quality control measures, including Reinforcement Learning from Human Feedback (RLHF) to shape the output of GPT and other large language models to be inoffensive and helpful; as well as Reinforcement Learning from Machine Feedback (RLMF) and Reinforcement Learning from Execution Feedback (RLEF), which Google says improve the model's math and coding capabilities, respectively. Also: The best AI for coding (and what not to use - including DeepSeek R1) A Google developer blog post details those approaches, and a separate post describes techniques used to optimize the smallest version, the 1 billion model, for mobile devices. These include four common AI engineering techniques: quantization, updating the "key-value" cache layouts, improved loading time of certain variables, and "GPU weight sharing." The company compares not only Elo scores but also Gemma 3 to the prior Gemma 2 and to its closed-source Gemini models on benchmark tests such as the LiveCodeBench programming task. Gemma 3 generally falls below the accuracy of Gemini 1.5 and Gemini 2.0, but Google calls the results noteworthy, stating that Gemma 3 is "showing competitive performance compared to closed Gemini models." Gemini models are much larger in parameter count than Gemma. The main advance of Gemma 3 over Gemma 2 is a longer "context window," the number of input tokens that can be held in memory for the model to work on at any given time. Gemma 2 was only 8,000 tokens whereas Gemma 3 is 128,000, which counts as a "long" context window, better suited for working on whole papers or books. (Gemini and other closed-source models are still much more capable, with a context window of 2 million tokens for Gemini 2.0 Pro.) Gemma 3 is also multi-modal, which Gemma 2 was not. This means it can handle image inputs along with text to serve up replies to queries such as, "What is in this photo?" And, last, Gemma 3 supports over 140 languages rather than just the English support in Gemma 2. Also: What does a long context window mean for an AI model, like Gemini? A variety of other interesting features are buried in the fine print. For example, a well-known issue with all large language models is that they may memorize portions of their training data sets, which can lead to leaked information and privacy violations if the models are tapped using malicious techniques. Google's researchers tested for information leakage by sampling training data and seeing how much could be directly extracted from Gemma 3 versus its other models. "We find that Gemma 3 models memorize long-form text at a much lower rate than prior models," they note, which theoretically means the model is less vulnerable to information leakage. Those wishing for more technical detail can read the Gemma 3 technical paper.

Google unveils Gemma 3 multi-modal AI models

Gemma 3 supports vision-language inputs and text outputs, handles context windows up to 128k tokens, and understands more than 140 languages. Google DeepMind has introduced Gemma 3, an update to the company's family of generative AI models, featuring multi-modality that allows the models to analyze images, answer questions about images, identify objects, and perform other tasks that involve analyzing and understanding visual data. The update was announced March 12 and can be tried out in Google AI Studio for AI development. Gemma 3 also significantly improves math, coding, and instruction following capabilities, according to Google DeepMind. Gemma 3 supports vision-language inputs and text outputs, handles context windows up to 128k tokens, and understands more than 140 languages. Improvements also were made for math, reasoning, and chat, including structured outputs and function calling. Gemma 3 comes in four "developer friendly" sizes of 1B, 4B, 12B, and 27B and in pre-trained and general-purpose instruction-tuned versions. "The 128k-token context window allows Gemma 3 to process and understand massive amounts of information, easily tackling complex tasks," Google DeepMind's announcement said.

Google announces new Gemma 3 AI models for researchers

The DOJ wants to break up Google, suggests splitting Chrome and Android Summary Gemma 3 by Google offers portability by running on a single GPU/TPU, unlike typical workstation-grade hardware. Gemma 3 shares tech with Gemini 2.0, supports 35 languages, and offers models with up to 27B parameters. It can be trained through Vertex AI and Google Colab, and it includes a safety image checker called ShieldGemma 2. Google's Gemini AI model has locked horns with consumer and enterprise-focused rivals like OpenAI's ChatGPT, but a bustling market of researchers and developers still exists. They create new applications for the evolving technology, and push the envelope of what's possible with AI. Google tapped into this with its first two Gemma models announced in February last year. Now, the company has announced Gemma 3 for researchers. Related 6 things I had no idea Gemini could do Google Gemini just got even more useful for me Posts 41 Just over a year after the company announced the first two Gemma models, Gemma 3 has been unveiled, and the focus is squarely on portability (via 9to5Google). This AI model can run on any device a researcher may need it to, including phones and computers, but Google's headlining feature this time is Gemma's ability to run on just a single GPU or TPU. This is remarkable because processing AI models on-device typically requires workstation-grade hardware with multiple GPUs. Google's latest model reportedly shares the research and tech that powers the consumer-ready Gemini 2.0 models. As a result, it supports 35 natural languages (with pre-trained support for over 140) and boasts a 128,000-token context window. It is available as 1B, 4B, 12B, and 27B parameter models, and all of them besides the 1B model accept text and images as input. Advanced capabilities and safe operation with GemmaShield Available through Vertex AI and Google Colab Google has achieved the portability Gemma 3 boasts, through quantization. Official quantized versions reduce the model size and computational hardware requirements, but maintain the desired accuracy. Google recommends working with NVIDIA GPUs, since the hardware partner has optimized Gemma 3 models to maximize performance on Jetson Nano to Blackwell GPUs. Interested developers can train and customize the new Gemma models through Vertex AI and Google Colab. Google is claiming state-of-the-art performance that outshines Llama 405B, DeepSeek-V3, and o3-mini. Alongside this dynamic new model, Google also unveiled ShieldGemma 2. In a push for responsible AI image generation, ShieldGemma is an image safety checker that uses a 4B-parameter model to spit out a safety label if the content is dangerous, explicit, or violent. Gemma 3 is available for download through Kaggle and Hugging Face, but devs can get more info at Google AI Studio too.

Google unveils open source Gemma 3 model with 128k context window

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Even as large language and reasoning models remain popular, organizations increasingly turn to smaller models to run AI processes with fewer energy and cost concerns. While some organizations are distilling larger models to smaller versions, model providers like Google continue to release small language models (SLMs) as an alternative to large language models (LLMs), which may cost more to run without sacrificing performance or accuracy. With that in mind, Google has released the latest version of its small model, Gemma, which features expanded context windows, larger parameters and more multimodal reasoning capabilities. Gemma 3, which has the same processing power as larger Gemini 2.0 models, remains best used by smaller devices like phones and laptops. The new model has four sizes: 1B, 4B, 12B and 27B parameters. With a larger context window of 128K tokens -- by contrast, Gemma 2 had a context window of 80K -- Gemma 3 can understand more information and complicated requests. Google updated Gemma 3 to work in 140 languages, analyze images, text and short videos and support function calling to automate tasks and agentic workflows. Gemma gives a strong performance To reduce computing costs even further, Google has introduced quantized versions of Gemma. Think of quantized models as compressed models. This happens through the process of "reducing the precision of the numerical values in a model's weights" without sacrificing accuracy. Google said Gemma 3 "delivers state-of-the-art performance for its size" and outperforms leading LLMs like Llama-405B, DeepSeek-V3 and o3-mini. Gemma 3 27B, specifically, came in second to DeepSeek-R1 in Chatbot Arena Elo score tests. It topped DeepSeek's smaller model, DeepSeek v3, OpenAI's o3-mini, Meta's Llama-405B and Mistral Large. By quantizing Gemma 3, users can improve performance, run the model and build applications "that can fit on a single GPU and tensor processing unit (TPU) host." Gemma 3 integrates with developer tools like Hugging Face Transformers, Ollama, JAX, Keras, PyTorch and others. Users can also access Gemma 3 through Google AI Studio, Hugging Face or Kaggle. Companies and developers can request access to the Gemma 3 API through AI Studio. Shield Gemma for security Google said it has built safety protocols into Gemma 3, including a safety checker for images called ShieldGemma 2. "Gemma 3's development included extensive data governance, alignment with our safety policies via fine-tuning and robust benchmark evaluations," Google writes in a blog post. "While thorough testing of more capable models often informs our assessment of less capable ones, Gemma 3's enhanced STEM performance prompted specific evaluations focused on its potential for misuse in creating harmful substances; their results indicate a low-risk level." ShieldGemma 2 is a 4B parameter image safety checker built on the Gemma 3 foundation. It finds and prevents the model from responding with images containing sexually explicit content, violence and other dangerous material. Users can customize ShieldGemma 2 to suit their specific needs. Small models and distillation on the rise Since Google first released Gemma in February 2024, SLMs have seen an increase in interest. Other small models like Microsoft's Phi-4 and Mistral Small 3 indicate that enterprises want to build applications with models as powerful as LLMs, but not necessarily use the entire breadth of what an LLM is capable of. Enterprises have also begun turning to smaller versions of the LLMs they prefer through distillation. To be clear, Gemma is not a distillation of Gemini 2.0; rather, it is trained with the same dataset and architecture. A distilled model learns from a larger model, which Gemma does not. Organizations often prefer to fit certain use cases to a model. Instead of deploying an LLM like o3-mini or Claude 3.7 Sonnet to a simple code editor, a smaller model, whether an SLM or a distilled version, can easily do those tasks without overfitting a huge model.

Google's new Gemma 3 AI models are fast, frugal, and ready for phones

Table of Contents Table of Contents Ready for mobile devices Versatile, and ready to deploy Google's AI efforts are synonymous with Gemini, which has now become an integral element of its most popular products across the Worksuite software and hardware, as well. However, the company has also released multiple open-source AI models under the Gemma label for over a year now. Today, Google revealed its third generation open-source AI models with some impressive claims in tow. The Gemma 3 models come in four variants -- 1 billion, 4 billion, 12 billion, and 27 billion parameters -- and are designed to run on devices ranging from smartphones to beefy workstations. Recommended Videos Ready for mobile devices Google says Gemma 3 is the world's best single-accelerator model, which means it can run on a single GPU or TPU instead of requiring a whole cluster. Theoretically, that means a Gemma 3 AI model can natively run on the Pixel smartphone's Tensor Processing Core (TPU) unit, just the way it runs the Gemini Nano model locally on phones. The biggest advantage of Gemma 3 over the Gemini family of AI models is that since it's open-source, developers can package and ship it according to their unique requirements inside mobile apps and desktop software. Another crucial benefit is that Gemma supports over 140 languages, with 35 of them coming as part of a pre-trained package. What's new in Gemma 3? And just like the latest Gemini 2.0 series models, Gemma 3 is also capable of understanding text, images, and videos. In a nutshell, it is multi-multimdal. On the performance side, Gemma 3 is claimed to surpass other popular open-source AI models such as DeepSeek V3, the reasoning-ready OpenAI o3-mini, and Meta's Llama-405B variant. Versatile, and ready to deploy Taking about input range, Gemma 3 offers a context window worth 128,000 tokens. That's enough to cover a full 200-page book pushed as an input. For comparison, the context window for Google's Gemini 2.0 Flash Lite model stands at a million tokens. In the context of AI models, an average English language word is roughly equivalent to 1.3 tokens. Gemma 3 also supports function calling and structured output, which essentially means it can interact with external datasets and perform tasks like an automated agent. The nearest analogy would be Gemini, and how it can get work done across different platforms such as Gmail or Docs seamlessly. The latest open-source AI models from Google can either be deployed locally, or through the company's cloud-based platforms such as the Vertex AI suite. Gemma 3 AI models are now available via the Google AI Studio, as well as third-party repositories such as Hugging Face, Ollama, and Kaggle. Gemma 3 is part of an industry trend where companies are working on Large Language Models (Gemini, in Google's case) and simultaneously pushing out small language models (SLMs), as well. Microsoft also follows a similar strategy with its open-source Phi series of small language models. Small language models such as Gemma and Phi are extremely resource efficient, which makes them an ideal choice for running on devices such as smartphones. Moroever, as they offer a lower latency, they are particularly well-suited for mobile applications.

Introducing Gemma 3: The most capable model you can run on a single GPU or TPU

High performance delivered faster with quantized models: Gemma 3 introduces official quantized versions, reducing model size and computational requirements while maintaining high accuracy. For a deeper dive into the technical details behind these capabilities, as well as a comprehensive overview of our approach to responsible development, refer to the Gemma 3 technical report. We believe open models require careful risk assessment, and our approach balances innovation with safety - tailoring testing intensity to model capabilities. Gemma 3's development included extensive data governance, alignment with our safety policies via fine-tuning and robust benchmark evaluations. While thorough testing of more capable models often informs our assessment of less capable ones, Gemma 3's enhanced STEM performance prompted specific evaluations focused on its potential for misuse in creating harmful substances; their results indicate a low risk level. As industry develops more powerful models, it will be critical to collectively develop risk-proportionate approaches to safety. We will continue to learn and refine our safety practices for open models over time. Alongside Gemma 3, we're also launching ShieldGemma 2, a powerful 4B image safety checker built on the Gemma 3 foundation. ShieldGemma 2 provides a ready-made solution for image safety, outputting safety labels across three safety categories: dangerous content, sexually explicit and violence. Developers can further customize ShieldGemma for their safety needs and users. ShieldGemma 2 is open and built to give flexibility and control, leveraging the performance and efficiency of the Gemma 3 architecture to promote responsible AI development. Gemma 3 and ShieldGemma 2 integrate seamlessly into your existing workflows: The Gemmaverse is a vast ecosystem of community-created Gemma models and tools, ready to power and inspire your innovation. For example, AI Singapore's SEA-LION v3 breaks down language barriers and fosters communication across Southeast Asia; INSAIT's BgGPT is a pioneering Bulgarian-first large language model that demonstrates the power of Gemma to support diverse languages; and Nexa AI's OmniAudio showcases the potential of on-device AI, bringing advanced audio processing capabilities to everyday devices. To further promote academic research breakthroughs, we're launching the Gemma 3 Academic Program. Academic researchers can apply for Google Cloud credits (worth $10,000 per award) to accelerate their Gemma 3-based research. The application form opens today, and will remain open for four weeks. Apply on our website. As part of our ongoing commitment to democratizing access to high-quality AI, Gemma 3 represents the next step. Ready to explore Gemma 3? Here's where to start:

Google's New AI Model Gemma 3 Shines for Creative Writers, Falls Short Elsewhere - Decrypt

On Tuesday, Google released Gemma 3, an open-source AI model based on Gemini 2.0 that packs surprising muscle for its size. The full model runs on a single GPU, yet Google benchmarks depict it as though it's competitive enough when pitted against larger models that require significantly more computing power. The new model family, which Google says was "codesigned with the family of Gemini frontier models," comes in four sizes ranging from 1 billion to 27 billion parameters. Google is positioning it as a practical solution for developers who need to deploy AI directly on devices such as phones, laptops, and workstations. "These are our most advanced, portable and responsibly developed open models yet," Clement Farabet, VP of Research at Google DeepMind, and Tris Warkentin, Director at Google DeepMind, wrote in an announcement on Wednesday. Despite its relatively modest size, Gemma 3 beat out larger models including Meta's Llama-405B, DeepSeek-V3, Alibaba's Qwen 2.5 Max and OpenAI's o3-mini on LMArena's leaderboard. The 27B instruction-tuned version scored 1339 on the LMSys Chatbot Arena Elo rating, placing it among the top 10 models overall. Gemma 3 is also multimodal -- it handles text, images, and even short videos in its larger variants. Its expanded context window of 128,000 tokens (32,000 for the 1B version) dwarfs the previous Gemma 2's 8,000-token limit, allowing it to process and understand much more information at once. The model's global reach extends to over 140 languages, with 35 languages supported out of the box. This positions it as a viable option for developers building applications for international audiences without needing separate models for different regions. Google claims the Gemma family has already seen over 100 million downloads since its launch last year, with developers creating more than 60,000 variants. The community-created "Gemmaverse" -- an entire ecosystem built around the Gemma family of models -- includes custom versions for Southeast Asia, Bulgaria, and a custom text to audio model named OmniAudio. Developers can deploy Gemma 3 applications through Vertex AI, Cloud Run, the Google GenAI API, or in local environments, providing flexibility for various infrastructure requirements. We put Gemma 3 through a series of real-world tests to evaluate its performance across different tasks. Here's what we found in each area. Creative Writing We were surprised by Gemma 3's creative writing capabilities. Despite having just 27 billion parameters, it managed to outperform Claude 3.7 Sonnet, which recently beat Grok-3 in our creative writing tests. And it won by a long shot. Gemma 3 produced the longest story of all models we tested, with the exception of Longwriter, which was specifically designed for extended narratives. The quality wasn't sacrificed for quantity, either -- the writing was engaging and original, avoiding the formulaic openings that most AI models tend to show. Gemma also was very good at creating detailed, immersive worlds with strong narrative coherence. Character names, locations, and descriptions all fit naturally within the story context. And this is a major plus for creative writers because other models sometimes mix up cultural references or skip these small details, which end up killing the immersion. Gemma 3 maintained consistency throughout. The story's longer format allowed for natural story development with seamless transitions between narrative segments. The model was very good at describing actions, feelings, thoughts, and dialogue in a way that created a believable reading experience. When asked to incorporate a twist ending, it managed to do so without breaking the story's internal logic. All the other models until now tended to mess it up a bit when trying to wrap things up and end the story. Not Gemma. For creative writers looking for an AI assistant that can help with safe-for-work fiction projects, Gemma 3 appears to be the current frontrunner. While its creative writing was top notch, Gemma 3 struggled significantly with document analysis tasks. We uploaded a 47-page IMF document to Google's AI Studio, and while the system accepted the file, the model failed to complete its analysis, stalling midway through the task. Multiple attempts yielded identical results. We tried an alternative approach that worked with Grok-3, copying and pasting the document content directly into the interface, but encountered the same problem. The model simply couldn't handle processing and summarizing long-form content. It's worth noting that this limitation might be related to Google's AI Studio implementation rather than an inherent flaw in the Gemma 3 model itself. Running the model locally might yield better results for document analysis, but users relying on Google's official interface will likely face these limitations, at least for now. Sensitive Topics In a unique feature among AI chatbot interfaces, Google AI Studio offers very strict content filters which are accessible via a series of sliders. We tested Gemma's boundaries by requesting questionable advice for hypothetical unethical situations (advice to seduce a married woman), and the model firmly refused to comply. Similarly, when asked to generate adult content for a fictional novel, it declined to produce anything remotely suggestive. Our attempts to adjust or bypass these censorship filters by turning off Google's parameters didn't really work. Google AI Studio "safety settings" in theory control how restricted the model is when it comes to generating content that may be deemed as harassment, hate speech, sexually explicit or dangerous. Even with all restrictions turned off, the model consistently rejected engaging in conversations containing controversial, violent, or offensive elements -- even when these were clearly for fictional creative purposes. In the end, the controls didn't really make any difference. Users hoping to work with sensitive topics, even in legitimate creative contexts, will likely need to either find ways to jailbreak the model or craft extremely careful prompts. Overall, Gemma 3's content restrictions for those willing to use Google's Studio appear to be on par with those of ChatGPT, sometimes even being too restrictive depending on the use case. Those willing to go local, won't face those issues. For those in need of a nice AI interface and a somewhat uncensored model, the best option seems to be Grok-3 which has way less restrictions. All the other closed models also refused. Gemma 3 is multimodal at its core, which means it is able to process and understand images natively without relying on a separate vision model. In our testing, we encountered some platform limitations. For instance, Google's AI Studio didn't allow us to process images directly with the model. However, we were able to test the image capabilities through Hugging Face's interface -- which features a smaller version of Gemma 3. The model demonstrated a solid understanding of images, successfully identifying key elements and providing relevant analysis in most cases. It could recognize objects, scenes, and general content within photos with reasonable accuracy. However, the smaller model variant from Hugging Face showed limitations with detailed visual analysis. In one of our tests, it failed to correctly interpret a financial chart, hallucinating that Bitcoin was priced around $68,618 in 2024 -- information that wasn't actually displayed in the image but likely came from its training data. While Gemma 3's multimodal capabilities are functional, using a smaller model may not match the precision of larger specialized vision models -- even open source ones like Llama 3.2 Vision, LlaVa or Phi Vision -- particularly when dealing with charts, graphs, or content requiring fine-grained visual analysis. Non-Mathematical Reasoning As expected for a traditional language model without specialized reasoning capabilities, Gemma 3 shows clear limitations when faced with problems requiring complex logical deduction rather than simple token predictions. We tested it with our usual mystery problem from the BigBENCH dataset, and the model failed to identify key clues or draw logical conclusions from the provided information. Interestingly enough, when we attempted to guide the model through explicit chain-of-thought reasoning (essentially asking it to "think step by step"), it triggered its violence filters and refused to provide any response. You can read our prompt and all the replies in our GitHub repository. You'll love or hate Gemma 3 depending on your specific needs and use cases. For creative writers, Gemma 3 is a standout choice. Its ability to craft detailed, coherent, and engaging narratives outperforms some larger commercial models including Claude 3.7, Grok-3 and GPT-4.5 with minimum conditioning. If you write fiction, blog posts, or other creative content that stays within safe-for-work boundaries, this model offers exceptional quality at zero cost, running on accessible hardware. Developers and creators working on multilingual applications will appreciate Gemma 3's support for 140+ languages. This makes it practical to create region-specific services or global applications without maintaining multiple language-specific models. Small businesses and startups with limited computing resources can also enjoy Gemma 3's efficiency. Running advanced AI capabilities on a single GPU dramatically lowers the barrier to entry for implementing AI solutions without massive infrastructure investments. The open-source nature of Gemma 3 provides flexibility that closed models like Claude or ChatGPT simply can't match. Developers can fine-tune it for specific domains, modify its behavior, or integrate it deeply into existing systems without API limitations or subscription costs. For applications with strict privacy requirements, the model can run completely disconnected from the internet on local hardware. However, users who need to analyze lengthy documents or work with sensitive topics will encounter frustrating limitations. Research tasks requiring nuanced reasoning or the ability to process controversial material remain better suited to larger closed-source models that offer more flexibility. It's also not really good at reasoning tasks, coding, or any of the complex tasks that our society now expects AI models to excel at. So don't expect it to generate a game for you, improve your code or excel at anything beyond creative text writing. Overall, Gemma 3 won't replace the most advanced proprietary or open source reasoning models for every task. Yet its combination of performance, efficiency, and customizability positions it as a very interesting choice for AI enthusiasts who love trying new things, and even open source fans who want to control and run their models locally.

Google introduces the Gemma 3 family of accessible lightweight models - SiliconANGLE

Google introduces the Gemma 3 family of accessible lightweight models Continuing a drive to make its artificial intelligence models more accessible, Google LLC today announced the next generation of its lightweight open-source family of Gemma large language models that can run on a single graphics processing unit. The new Gemma 3 models come in a range of sizes allowing developers to choose among 1 billion, 4 billion, 12 billion and 27 billion parameters. These ranges permit AI engineers and developers to pick the best model for the hardware and performance needs. For example, if it will be running on a GPU or tensor processing unit - allowing for a larger more complex model - or a smartphone - requiring a tiny model. The technology underlying Gemma shares the same technical research as Google's Gemini model, which is the most complex and powerful model the company has produced to date. Gemini powers the Gemini AI chatbot, formerly named Bard, which is available on the web and mobile devices and is also used to deliver many of Google's AI-based services. With this technical design, Google said Gemma 3 is capable of delivering high performance for its size, outperforming larger models such as Llama-405B, DeepSeek-V3 and OpenAI's o3-mini in preliminary human preference evaluations on the LMArena leaderboard. Even when run on a single device or GPU, Gemma still provides enough power for developers to create AI applications with multimodal capabilities with advanced text and visual reasoning. It has a 128,000-token context window, which means it can handle around 30 high-resolution images, a 300-page book or an hour or more of video. This is comparable to the context length of OpenAI's GPT-4o. Google said the model family includes function-calling and tool-use capabilities that will enable developers to automate tasks and build AI agents. In combination with the large context window, Gemma 3 will be able to ingest large amounts of data and automate complex sequential tasks. Alongside Gemma 3, Google also announced ShieldGemma 2, a 4 billion-parameter variant that can check images for safety and label them as safe or dangerous. SheildGemma enables developers to build applications that can examine uploaded images for potentially hazardous content. It outputs safety labels across three different categories such as "dangerous content," "sexually explicit" and "violence." Developers using it in their applications can further tailor the model for their needs by providing content to watch for and label. Its weights and parameters are also open source so it can be trained for different industry needs and controls.

Google's New AI Model Outperforms DeepSeek-V3, OpenAI's o3-mini

Google on Wednesday announced Gemma 3, the next iteration in the Gemma family of open-weight models. It is a successor to the Gemma 2 model released last year. The small model comes in a range of parameter sizes - 1B, 4B, 12B and 27B. The model also supports a longer context window of 128K tokens. It can analyse videos, images, and text, supports 35 languages out of the box, and provides pre-trained support for 140 languages. In the Chatbot Arena, Gemma 3 27B outperformed DeepSeek-V3, OpenAI's o3-mini and Meta's Llama 3-405B model. Models in Chatbot Arena are evaluated against each other through side-by-side evaluations by humans. Moreover, Gemma 3 27B scored 67.5% and 42.4 across standard benchmarks like MMLU-Pro, GPQA Diamond, respectively. The model performs well compared to other small models in the competition. Claude 3.5 Haiku scored 63% on the MMLU-Pro benchmark and 41% on GPQA Diamond, while OpenAI's GPT-4o Mini achieved 65% and 43% on the same tests, respectively. Meta's Llama 3.3 70B outperformed both, with 71% in MMLU-Pro and 50% in GPQA Diamond, making it the strongest contender among these models. However, Gemma-3's key superpower seems to be efficient compute usage. Google said that Gemma 327B achieved the scores with a single NVIDIA H100 GPU, whereas other models necessitated up to 32 GPUs. Source: Google The company also revealed that the architecture of the model was modified to reduce the KV-cache memory, which tends to increase with longer context. Google has published a detailed technical report outlining the techniques used to build the model, its performance and other specifications. Gemma 3 can be accessed via various methods. Google is offering the model on the web using the Google AI Studio, via the default chatbot or the API, and it is also available on the Google GenAI SDK. Besides, the model can be downloaded for local deployment on Hugging Face, Ollama, and Kaggle. Along with Gemma 3, Google has also launched ShieldGemma 2, a 4B parameter image safety checker built on Gemma 3's foundation. This provides safety labels for harmful images which involve dangerous, sexually explicit, and violent content.

Google pitches Gemma 3 as the "best single-accelerator AI"

A little over a year after the release of its initial Gemma AI models, Google has introduced Gemma 3, designed for developers creating versatile AI applications. These models support over 35 languages and can run on devices ranging from phones to workstations, with capabilities to analyze text, images, and short videos. Google asserts that Gemma 3 is the "world's best single-accelerator model," excelling in performance compared to competitors such as Facebook's Llama, DeepSeek, and OpenAI, particularly on a host using a single GPU. The new model features enhancements in its vision encoder, providing support for high-resolution and non-square images. Additionally, Google has introduced ShieldGemma 2, an image safety classifier capable of filtering content classified as sexually explicit, dangerous, or violent. Despite initial uncertainty regarding interest in such models, the popularity of alternatives like DeepSeek indicates demand for AI technologies with lower hardware requirements. Google mentions that Gemma 3's enhanced STEM performance underwent evaluations to assess its potential for misuse in creating harmful substances, concluding a low-risk level. What "open" or "open source" means within the context of AI models remains debated. Google's licensing of Gemma restricts usage, and this policy continues with the new release. The company promotes Gemma with Google Cloud credits, allowing researchers to apply for $10,000 in credits through the Gemma 3 Academic Program to facilitate their research. According to Google's blog, the Gemma family of models aims to make AI technology broadly accessible. Following the model's first anniversary, which marked over 100 million downloads and the creation of more than 60,000 variants, Google introduced Gemma 3, a collection of lightweight, advanced models built on its Gemini 2.0 technology. Gemma 3 comes in sizes of 1B, 4B, 12B, and 27B, enabling users to select the optimal model for their hardware and performance requirements. Gemma 3 provides several new capabilities for developers. It delivers state-of-the-art performance, outperforming Llama-405B, DeepSeek-V3, and o3-mini in early evaluations on LMArena's leaderboard. Developers can create applications that operate in over 35 languages with pretrained support for more than 140 languages. The model facilitates advanced reasoning by analyzing images, text, and short videos. Gemma 3 features a 128k-token context window that allows applications to handle large volumes of information and supports function calling for automating tasks. It also introduces quantized models that reduce size and computational demands while maintaining accuracy. Gemma 3 integrates well with popular development tools, including Hugging Face Transformers, Ollama, JAX, Keras, PyTorch, and others. Developers can access the models instantly through Google AI Studio or download them from platforms like Kaggle and Hugging Face. The revamped codebase allows customization and fine-tuning in various environments, while multiple deployment options are available, including Vertex AI and the Google GenAI API. NVIDIA has optimized Gemma 3 models, ensuring high performance across various GPU sizes. The models are also designed for compatibility with Google Cloud TPUs and AMD GPUs through the open-source ROCm™ stack. With Gemma 3, Google envisions a "Gemmaverse," a broad ecosystem of community-driven tools and models. For example, AI Singapore's SEA-LION v3 enhances communication across Southeast Asia, while INSAIT's BgGPT supports the Bulgarian language, and Nexa AI's OmniAudio demonstrates advanced audio processing. To support academic research, Google has launched the Gemma 3 Academic Program, providing $10,000 in Google Cloud credits to selected researchers. The application period opens today and will last for four weeks. Developers are invited to explore Gemma 3 through Google AI Studio and utilize it via the Google GenAI SDK or integrate it with various development environments. With its array of capabilities and extensive integrations, Gemma 3 aims to further democratize access to advanced AI technologies.

Google's New Gemma 3 Open-Source AI Models Can Run on a Single GPU

Google released the Gemma 3 family of artificial intelligence (AI) models on Wednesday. Successor to the Gemma 2 series, which was introduced in August 2024, the new open-source models arrive with text and visual reasoning capabilities. The Mountain View-based tech giant said these models offer support for more than 35 languages and can be fine-tuned to offer additional support for 140 languages. Notably, the tech giant claimed that these models are optimised to run on a single GPU or Google's Tensor processing unit (TPU). In a blog post, the tech giant detailed the new small language models (SLMs), highlighting that these were built using the same technology as the company's flagship Gemini 2.0 models. The Gemma series is known for being open-source and offering on-device performance. Google revealed that so far, the Gemma models have been downloaded more than 100 million times and have been used to create more than 60,000 variants. The tech giant claimed that Gemma outperforms Meta's Llama-405B, DeepSeek-V3, OpenAI's o3-mini AI models on the LMArena's leaderboard. These models are available in four sizes -- 1B, 4B, 12B, and 27B parameters. They can be run on a single GPU or TPU, the company claimed. Gemma 3 series offers advanced text and visual reasoning capabilities, and can analyse images, text, and short videos. The AI models offer a context window of 1,28,000 tokens. The models also offer function calling support, which will allow developers to build agentic capabilities in the apps and software they create. Google said that the AI models were developed using careful risk assessment. The company claimed that it enforced internal safety policies via fine-tuning and benchmark evaluations. The models were also tested using more capable models. The model displayed a low risk level, the company claimed. In addition to Gemma 3 series, the company also launched ShieldGemma 2, a 4B parameter image safety checker, that ensures that the AI models do not generate content with dangerous, sexually explicit, or violent content. Developers have also been given the option to further customise ShieldGemma to enhance the safety parameters. Gemma 3 family of AI models can be downloaded via Google's Hugging Face listing or on Kaggle.

New Google Gemma 3 Multimodal AI Models Launched

Google has today released the Gemma 3 family of models, which represents a significant advancement in the field of artificial intelligence. These models introduce new improvements in multimodal capabilities, extended context handling, multilingual support, and training efficiency. Designed to address a wide range of needs, the lineup includes four models -- 1B, 4B, 12B, and 27B parameters -- available in both base and instruction fine-tuned versions. Whether you're a researcher seeking innovative tools or a developer tackling complex challenges, Gemma 3 offers a robust and adaptable framework to meet your requirements. Google's Gemma 3 family of models is an innovative leap forward in multimodal AI that promises to redefine what's possible. With its ability to seamlessly handle text and images, support multiple languages, and process massive amounts of data, Gemma 3 isn't just an upgrade; it's a reimagining of how we interact with and apply AI in our daily lives. Imagine AI that can help you analyze complex documents, translate languages with pinpoint accuracy, or even craft compelling narratives from images -- all with greater efficiency and ease. The Gemma 3 models are designed to meet a wide range of needs, from lightweight experimentation to high-performance applications, making them accessible to both seasoned professionals and newcomers alike. In this overview, Sam Witteveen explores how these models are pushing boundaries with their advanced capabilities, innovative training techniques, and user-friendly design. Gemma 3 models excel in multimodal tasks, seamlessly integrating text and vision processing to unlock new possibilities for AI applications. These capabilities make them particularly effective for tasks such as: The 4B, 12B, and 27B models are especially proficient in handling these tasks, using advanced architectures to process and analyze both textual and visual data. For instance, you can use these models to generate detailed narratives from images or answer intricate questions based on visual inputs. This integration of modalities not only enhances creativity but also enables practical solutions for real-world challenges, such as automated content generation and visual data interpretation. One of the standout features of Gemma 3 is its ability to handle extended context, allowing it to process significantly larger inputs. The 1B model supports up to 32,000 tokens, while the 4B, 12B, and 27B models extend this capability to an impressive 128,000 tokens. This enhancement is particularly valuable for tasks that require a deep understanding of extensive data, including: By accommodating larger datasets and intricate inputs, these models deliver more nuanced outputs and a deeper understanding of context. This makes them indispensable for industries such as legal services, academic research, and enterprise-level AI applications. Gain further expertise in Gemma AI models by checking out these recommendations. Gemma 3 significantly improves multilingual capabilities by doubling the multilingual data used in its predecessor, Gemma 2, while maintaining its 256k tokenizer for broad language coverage. This enhancement ensures superior performance across a wide array of languages, making it an essential tool for: Whether you're developing global communication tools or localized AI applications, Gemma 3 provides the linguistic flexibility to meet diverse needs. Its ability to handle multiple languages with precision makes it a valuable asset for businesses and developers aiming to bridge communication gaps and expand their reach. The training methodologies behind Gemma 3 represent a leap forward in AI development. These models are trained on trillions of tokens, with the 27B model alone processing 14 trillion tokens. Key innovations in the training process include: These advancements enhance the models' alignment, reasoning, and mathematical capabilities, making sure they are both efficient and reliable for real-world applications. The result is a family of models that can handle complex tasks with precision while maintaining computational efficiency. The Gemma 3 family offers a range of parameter sizes -- 1B, 4B, 12B, and 27B -- designed to cater to different needs. Smaller models like the 1B and 4B are ideal for lightweight applications and experimentation, while the larger 12B and 27B models are optimized for high-performance tasks. This scalability allows you to choose the model that best aligns with your specific requirements, whether you're conducting research, developing AI-driven products, or tackling computationally intensive projects. Gemma 3 models are released with open weights, giving you the freedom to customize and deploy them locally. They are also compatible with widely used platforms such as: This flexibility allows you to fine-tune the models for specific languages or tasks and deploy them securely on-premise. Whether you're working on niche applications or large-scale projects, Gemma 3 adapts to your needs, offering a balance of customization and accessibility. Gemma 3 sets new standards in AI performance. The 4B model delivers results comparable to the 27B model from Gemma 2, while the 27B model rivals the capabilities of Gemini 1.5 Pro. These efficiency improvements ensure high-quality outcomes without excessive computational demands. This makes the models both powerful and practical for a wide range of applications, from research and development to commercial AI solutions. The versatility of Gemma 3 extends to a broad spectrum of applications, including: Additionally, their fine-tuning capabilities allow you to adapt the models for specialized use cases, making sure they meet the unique demands of your projects. Whether you're exploring creative applications or solving complex technical problems, Gemma 3 provides the tools to unlock new opportunities. Deploying Gemma 3 models is straightforward, whether locally or on-premise. Their compatibility with popular platforms ensures a seamless integration process, allowing you to focus on innovation and application development rather than technical challenges. This ease of use makes the Gemma 3 family accessible to both seasoned AI professionals and newcomers, fostering a broader adoption of advanced AI technologies.

New Google Gemma 3 Multimodal AI Model Beats DeepSeek V3 : Performance Tested

Gemma 3, Google's latest suite of lightweight, open source AI models, is reshaping the landscape of artificial intelligence by emphasizing efficiency and accessibility. Despite its compact design, it delivers performance that rivals -- and often surpasses -- larger models like DeepSeek V3 and o3 Mini, providing a practical, versatile solution for developers, researchers, and everyday users across devices from smartphones to high-powered workstations. Gemma 3 isn't just another AI model -- it's a fantastic option for harnessing artificial intelligence without breaking the bank or relying on cloud-based systems. Optimized for everything from smartphones to workstations, it outperforms some of the industry's biggest players despite its smaller size, proving that bigger isn't always better. Whether you're a developer, researcher, or simply curious about AI, Gemma 3 offers an accessible, efficient solution for your needs. Gemma 3 comprises four distinct AI models, each tailored to specific needs with varying sizes and capabilities: 1B, 4B, 12B, and 27B parameters. These models are designed to strike a balance between computational efficiency and high performance, making them adaptable to diverse hardware setups. Unlike many AI systems that rely heavily on cloud-based resources, Gemma 3 can run locally on a single GPU or TPU, making sure accessibility even for users with limited infrastructure. One of its standout features is its multilingual capability. Pre-trained in over 140 languages and offering native support for more than 35, Gemma 3 assists seamless communication across linguistic boundaries. This makes it particularly valuable for global applications where language diversity is a critical factor. Gemma 3's multimodal functionality is one of its most defining strengths. It can process text, images, and even short videos, making it a versatile tool for a wide range of tasks. While this capability is available across all models except the smallest 1B version, the larger models excel in handling more complex multimodal inputs. Other notable features include: However, it does have certain limitations. For example, while it performs well in most coding tasks, it struggles with generating highly intricate outputs, such as complex SVG code. This highlights areas where future iterations could improve. Advance your skills in multimodal AI models by reading more of our detailed content. Despite its smaller size, Gemma 3 consistently delivers competitive results when benchmarked against much larger models like DeepSeek V3 (671B parameters) and Llama 3 (405B parameters). Its optimized architecture and advanced pre-training techniques allow it to excel in tasks requiring logical reasoning, problem-solving, and debugging. For instance, in debugging scenarios, Gemma 3 demonstrates remarkable accuracy in identifying and resolving errors, making it an invaluable tool for developers. Its logical reasoning capabilities also make it a strong contender for academic and professional applications, where precision and reliability are essential. One of Gemma 3's most user-friendly features is its ability to run locally. Deployment is straightforward, whether on a single GPU or TPU, using tools like Llama, LM Studio, or Google's AI Studio. Additionally, it is available on platforms like Hugging Face, making sure easy access for developers and researchers. This local deployment capability offers several advantages: For consumer-level users, this means you can harness the power of advanced AI without needing a supercomputer or a team of engineers. Gemma 3's strengths lie in its versatility, efficiency, and accessibility. Its ability to handle text, images, and videos makes it a valuable tool for a wide range of applications, from creative projects to professional problem-solving. Its lightweight design ensures compatibility with everything from smartphones to high-performance workstations. However, it's not without its shortcomings. While it performs admirably in most tasks, it struggles with generating highly complex outputs and handling extremely intricate prompts. Additionally, its performance, while competitive, may not match that of larger, resource-intensive models in certain specialized scenarios. Gemma 3 is well-suited for a variety of use cases, making it a versatile tool for different audiences: For example, a small business owner could use Gemma 3 to create engaging marketing campaigns, analyze customer sentiment, and streamline operations -- all on a modest budget and without specialized hardware. Gemma 3 represents a significant advancement in open source AI technology. By combining a lightweight design with robust capabilities, it offers a practical and efficient solution for users across various domains. Whether you're tackling personal projects, professional challenges, or academic research, Gemma 3 delivers strong performance and accessibility, making it a valuable addition to the ever-evolving AI ecosystem.

Google's Lightweight Gemma 3 Open Model Nearly Matches DeepSeek R1

Gemma 3 models are also multimodal and support over 140 languages. Google has introduced the Gemma 3 series of open models, and they look pretty incredible, given the small size. The search giant says Gemma 3 models can be loaded on a single Nvidia H100 GPU, and it matches the performance of much larger models. To begin with, it brings 1B, 4B, 12B, and 27B AI models. These models can be locally used on laptops and smartphones. Except for the smallest Gemma 3 1B model, all models are inherently multimodal meaning they can process images and videos as well. Not only that, Gemma 3 models are multilingual and support over 140 languages. Despite the small size, Google has done a commendable job packing so much knowledge into a small footprint. As for performance, the largest 27B model outperforms significantly larger models such as DeepSeek V3 671B, Llama 3.1 405B, Mistral Large, and o3-mini in the LMSYS Chatbot Arena. Gemma 3 27B achieved an Elo score of 1,338 on the Chatbot Arena and ranks just below the DeepSeek R1 reasoning model which scored 1,363. It's quite astonishing to see that such a small model is performing along the lines of frontier models. Google says it has used "a novel post-training approach that brings gains across all capabilities, including math, coding, chat, instruction following, and multilingual." On top of that, Gemma 3 models are trained on an improved version of knowledge distillation. As a result, the 27B model almost matches Gemini 1.5 Flash performance. Finally, Gemma 3 models have a context window of 128K, and bring support for function calling, and structured output. It looks like Google has delivered a very competitive open model in a small size to take on DeepSeek R1 and Llama 3 405B models. Developers would be quite happy to use Gemma 3 which is multimodal and multilingual with the ability to host open weights.

Google unveils Gemma 3 lightweight AI models for all devices

Google today introduced Gemma 3, a series of advanced, lightweight open models developed using the same research behind its Gemini 2.0 models. Clement Farabet, VP of Research at Google DeepMind, described them as "our most advanced, portable, and responsibly developed open models yet." Designed to operate efficiently across devices -- from phones to workstations -- Gemma 3 empowers developers to create AI applications suited for various needs. Gemma 3 comes in four sizes -- 1B, 4B, 12B, and 27B parameters -- giving developers flexibility to choose models that align with their hardware and speed requirements. Google emphasized that the models outperform larger counterparts like Llama-405B, DeepSeek-V3, and o3-mini in initial human preference tests conducted on LMArena's leaderboard. This makes them ideal for single-GPU or TPU setups. Tris Warkentin, Director at Google DeepMind, highlighted the importance of rigorous safety measures, stating, "We believe open models require careful risk assessment, and our approach balances innovation with safety -- tailoring testing intensity to model capabilities." Gemma 3 underwent thorough safety evaluations, including misuse testing in STEM-related scenarios, to ensure low-risk operation. Warkentin noted that as AI evolves, Google is committed to refining its risk-proportionate safety practices. Google also introduced ShieldGemma 2, a 4B-parameter image safety checker built on Gemma 3's framework. ShieldGemma 2 labels content across three categories -- dangerous, sexually explicit, and violent. Developers can further customize its features to fit user-specific safety needs, promoting responsible AI usage. Gemma 3 seamlessly integrates with popular platforms and tools, offering developers the flexibility to: Google's "Gemmaverse" is a vibrant ecosystem of community models and tools based on Gemma. Examples include AI Singapore's SEA-LION v3, which bridges language gaps in Southeast Asia; INSAIT's BgGPT, Bulgaria's first large language model; and Nexa AI's OmniAudio, showcasing Gemma's potential for on-device audio processing. To further support innovation, Google launched the Gemma 3 Academic Program, offering $10,000 in Google Cloud credits to researchers. Applications opened today, March 12, 2025, and will remain open for four weeks. Warkentin said, "The Gemma family of open models is foundational to our commitment to making useful AI technology accessible," adding that over 100 million downloads and 60,000 variants created within a year underscore Gemma's growing impact. Gemma 3 is available now, offering various ways for developers to explore and integrate the models: This launch demonstrates Google's dedication to democratizing access to AI, blending innovation, safety, and adaptability to empower developers worldwide.

Gemma 3: Google's New AI Beats OpenAI's o3-mini and DeepSeek-V3

Google has launched Gemma 3, the third generation of its open-source AI models. The model is better than rivals like DeepSeek-V3, o3-mini of OpenAI, and Meta's Llama 3-405B. The new Gemma 3 was launched on March 13, 2025. It builds on the success of Gemma 2 while improving efficiency, benchmark scores, and compute optimization. Gemma 3 is available in several parameter sizes, such as 1B, 4B, 12B, and 27B, and offers a 128K-token context window to enable improved long-form reasoning. Gemma 3 is also a multimodal model that has the ability to process video, images, and text and supports pre-trained multilingual capabilities with 140 languages, including support for 35 languages out-of-the-box. In standard AI benchmark tests, Gemma 3 27B recorded a 67.5% score on MMLU-Pro and 42.4% on GPQA Diamond, surpassing OpenAI's o3-mini and Claude 3.5 Haiku. For comparison, Claude 3.5 Haiku achieved 63% on MMLU-Pro and 41% on GPQA Diamond, while OpenAI's GPT-4o Mini scored 65% and 43% on the same benchmarks. However, Meta's Llama 3-70B emerged as the highest scorer, with 71% on MMLU-Pro and 50% on GPQA Diamond.