Gemma is a family of lightweight, state-of-the art open models built from the same research and technology used to create the Gemini models.
The Keras team is happy to announce that Gemma, a family of lightweight, state-of-the art open models built from the same research and technology that we used to create the Gemini models, is now available in the KerasNLP collection. Thanks to Keras 3, Gemma runs on JAX, PyTorch and TensorFlow. With this release, Keras is also introducing several new features specifically designed for large language models: a new LoRA API (Low Rank Adaptation) and large scale model-parallel training capabilities.
If you want to dive directly into code samples, head here:
- Get started with Gemma models
- Fine-tune Gemma models with LoRA
- Fine-tune Gemma models on multiple GPUs/TPUs
Gemma models come in portable 2B and 7B parameter sizes, and deliver significant advances against similar open models, and even some larger ones. For example:
- Gemma 7B scores a new best-in class 64.3% of correct answers in the MMLU language understanding benchmark (vs. 62.5% for Mistral-7B and 54.8% for Llama2-13B)
- Gemma adds +11 percentage points to the GSM8K benchmark score for grade-school math problems (46.4% for Gemma 7B vs. Mistral-7B 35.4%, Llama2-13B 28.7%)
- and +6.1 percentage points of correct answers in HumanEval, a coding challenge (32.3% for Gemma 7B, vs. Mistral 7B 26.2%, Llama2 13B 18.3%).
Gemma models are offered with a familiar KerasNLP API and a super-readable Keras implementation. You can instantiate the model with a single line of code:
Posted by gemma_lm = keras_nlp.models.GemmaCausalLM.from_preset("gemma_2b_en")
And run it directly on a text prompt – yes, tokenization is built-in, although you can easily split it out if needed - read the Keras NLP guide to see how.
gemma_lm.generate("Keras is a", max_length=32)
> "Keras is a popular deep learning framework for neural networks..."
Try it out here: Get started with Gemma models
Thanks to Keras 3, you can choose the backend on which you run the model. Here is how to switch:
os.environ["KERAS_BACKEND"] = "jax" # Or "tensorflow" or "torch".
import keras # import keras after having selected the backend
Keras 3 comes with several new features specifically for large language models. Chief among them is a new LoRA API (Low Rank Adaptation) for parameter-efficient fine-tuning. Here is how to activate it:
gemma_lm.backbone.enable_lora(rank=4)
# Note: rank=4 replaces the weights matrix of relevant layers with the
# product AxB of two matrices of rank 4, which reduces the number of
# trainable parameters.
This single line drops the number of trainable parameters from 2.5 billion to 1.3 million!
Try it out here: Fine-tune Gemma models with LoRA.
Keras 3 also supports large-scale model training and Gemma is the perfect model to try it out. The new Keras distribution API offers data-parallel and model-parallel distributed training options. The new API is meant to be multi-backend but for the time being, it is implemented for the JAX backend only, because of its proven scalability (Gemma models were trained with JAX).
To fine-tune the larger Gemma 7B, a distributed setup is useful, for example a TPUv3 with 8 TPU cores that you can get for free on Kaggle, or an 8-GPU machine from Google Cloud. Here is how to configure the model for distributed training, using model parallelism:
device_mesh = keras.distribution.DeviceMesh(
(1, 8), # Mesh topology
["batch", "model"], # named mesh axes
devices=keras.distribution.list_devices() # actual accelerators
)
# Model config
layout_map = keras.distribution.LayoutMap(device_mesh)
layout_map["token_embedding/embeddings"] = (None, "model")
layout_map["decoder_block.*attention.*(query|key|value).*kernel"] = (
None, "model", None)
layout_map["decoder_block.*attention_output.*kernel"] = (
None, None, "model")
layout_map["decoder_block.*ffw_gating.*kernel"] = ("model", None)
layout_map["decoder_block.*ffw_linear.*kernel"] = (None, "model")
# Set the model config and load the model
model_parallel = keras.distribution.ModelParallel(
device_mesh, layout_map, batch_dim_name="batch")
keras.distribution.set_distribution(model_parallel)
gemma_lm = keras_nlp.models.GemmaCausalLM.from_preset("gemma_7b_en")
# Ready: you can now train with model.fit() or generate text with generate()
What this code snippet does is set up the 8 accelerators into a 1 x 8 matrix where the two dimensions are called “batch” and “model”. Model weights are sharded on the “model” dimension, here split between the 8 accelerators, while data batches are not partitioned since the “batch” dimension is 1.
Try it out here: Fine-tune Gemma models on multiple GPUs/TPUs.
We will soon be publishing a guide showing you how to correctly partition a Transformer model and write the 6 lines of partitioning setup above. It is not very long but it would not fit in this post.
You will have noticed that layer partitionings are defined through regexes on layer names. You can check layer names with this code snippet. We ran this to construct the LayoutMap above.
# This is for the first Transformer block only,
# but they all have the same structure
tlayer = gemma_lm.backbone.get_layer('decoder_block_0')
for variable in tlayer.weights:
print(f'{variable.path:<58} {str(variable.shape):<16}')
Full GSPMD model parallelism works here with just a few partitioning hints because Keras passes these settings to the powerful XLA compiler which figures out all the other details of the distributed computation.
We hope you will enjoy playing with Gemma models. Here is also an instruction-tuning tutorial that you might find useful. And by the way, if you want to share your fine-tuned weights with the community, the Kaggle Model Hub now supports user-tuned weights uploads. Head to the model page for Gemma models on Kaggle and see what others have already created!
Google has long believed that open technology is not only good for our company, but good for the industry, consumers, and the world. We’ve released open-source projects like Android and Chromium that transformed access to mobile and web technologies, and have done the same in AI with Transformers, TensorFlow, and AlphaFold. The release of our Gemma family of open models is a next step in how we’re deepening our commitment to open technology alongside an industry-leading safe, responsible approach. At the same time, the rapidly evolving nature of AI raises important considerations for how to enable safety-aligned open models: an approach that supports broad innovation while promoting safe uses.
A benefit of open source is that once it is released, its license gives users full creative autonomy. This is a powerful guarantee of technology access for developers and end users. Another benefit is that open-source technology can be modified to fit the unique use case of the end user, without restriction.
In the hands of a malicious actor, however, the lack of restrictions can raise risks. Computing has been through similar cycles before, addressing issues such as protecting users of the open internet, handling cryptography, and addressing open-source software security. We now face this challenge with AI. Below we share the approach we took to openly releasing Gemma models, and the advancements in open model safety we hope to accelerate.
Today, Gemma models are being released as what the industry collectively has begun to refer to as “open models.” Open models feature free access to the model weights, but terms of use, redistribution, and variant ownership vary according to a model’s specific terms of use, which may not be based on an open-source license. The Gemma models’ terms of use make them freely available for individual developers, researchers, and commercial users for access and redistribution. Users are also free to create and publish model variants. In using Gemma models, developers agree to avoid harmful uses, reflecting our commitment to developing AI responsibly while increasing access to this technology.
We’re precise about the language we’re using to describe Gemma models because we’re proud to enable responsible AI access and innovation, and we’re equally proud supporters of open source. The definition of "Open Source" has been invaluable to computing and innovation because of requirements for redistribution and derived works, and against discrimination. These requirements enable cross-industry collaboration, individual innovation and entrepreneurship, and shared research to happen with exponential effects.
However, existing open-source concepts can’t always be directly applied to AI systems, which raises questions on how to use open-source licenses with AI. It’s important that we carry forward open principles that have made the sea-change we’re experiencing with AI possible while clarifying the concept of open-source AI and addressing concepts like derived work and author attribution.
Licensing and terms of use are only one part of the evaluations, technical tools, and considered decision-making that went into aligning this release with our responsible AI Principles. Our approach involved:
We continue to evolve our approach. As we build these frameworks further, we will proceed thoughtfully and incorporate what we learn into future model assessments. We will continue to explore the full range of access mechanisms, with benefits and risk mitigation in mind, including API-based access and staged releases.
Many of today’s AI safety tools are designed for systems where the design approach assumes restricted access and redistribution, as well as auxiliary controls like query filters. Similarly, much of the AI safety research for improving mitigations takes on the design assumptions of those systems. Just as we have created unique threat models and solutions for other open technology, we are developing safety and security tools appropriate for the differences of openly available AI.
As models become more and more capable, we are conducting research and investing in rigorous safety evaluation, testing, and mitigations for open models. We are also actively participating in conversations with policymakers and open-source community leaders on how the industry should approach this technology. This challenge is multifaceted, just like AI systems themselves. Model-sharing platforms like Hugging Face and Kaggle, where developers inspire each other with novel model iterations, play a critical role in efforts to develop open models safely; there is also a role for the cybersecurity community to contribute learnings and best practices.
Building those solutions requires access to open models, sharing innovations and improvements. We believe sharing the Gemma models will not just help increase access to AI technology, but also help the industry develop new approaches to safety and responsibility.
As developers adopt Gemma models and other safety-aligned open models, we look forward to working with the open-source community to develop more solutions for responsible approaches to AI in the open ecosystem. A global diversity of experiences, perspectives, and opportunities will help build safe and responsible AI that works for everyone.
By Anne Bertucio – Sr Program Manager, Open Source Programs Office; Helen King – Sr Director of Responsibility, Google DeepMind
Gemini 1.5 Pro brings big improvements to speed and efficiency, but one of its innovations is its long context window, which measures how many tokens that the model can …
For 60 years, the Munich Security Conference has brought together world leaders, businesses, experts and civil society for frank discussions about strengthening and safe…
Today we are open-sourcing Magika, Google’s AI-powered file-type identification system, to help others accurately detect binary and textual file types. Under the hood, Magika employs a custom, highly optimized deep-learning model, enabling precise file identification within milliseconds, even when running on a CPU.
 |
| Magika command line tool used to recognize a identify the type of a diverse set of files |
You can try the Magika web demo today, or install it as a Python library and standalone command line tool (output is showcased above) by using the standard command line pip install magika.
Since the early days of computing, accurately detecting file types has been crucial in determining how to process files. Linux comes equipped with libmagic and the file utility, which have served as the de facto standard for file type identification for over 50 years. Today web browsers, code editors, and countless other software rely on file-type detection to decide how to properly render a file. For example, modern code editors use file-type detection to choose which syntax coloring scheme to use as the developer starts typing in a new file.
Accurate file-type detection is a notoriously difficult problem because each file format has a different structure, or no structure at all. This is particularly challenging for textual formats and programming languages as they have very similar constructs. So far, libmagic and most other file-type-identification software have been relying on a handcrafted collection of heuristics and custom rules to detect each file format.
This manual approach is both time consuming and error prone as it is hard for humans to create generalized rules by hand. In particular for security applications, creating dependable detection is especially challenging as attackers are constantly attempting to confuse detection with adversarially-crafted payloads.
To address this issue and provide fast and accurate file-type detection we researched and developed Magika, a new AI powered file type detector. Under the hood, Magika uses a custom, highly optimized deep-learning model designed and trained using Keras that only weighs about 1MB. At inference time Magika uses Onnx as an inference engine to ensure files are identified in a matter of milliseconds, almost as fast as a non-AI tool even on CPU.
 |
| Magika detection quality compared to other tools on our 1M files benchmark |
Performance wise, Magika, thanks to its AI model and large training dataset, is able to outperform other existing tools by about 20% when evaluated on a 1M files benchmark that encompasses over 100 file types. Breaking down by file type, as reported in the table below, we see even greater performance gains on textual files, including code files and configuration files that other tools can struggle with.
 |
| Various file type identification tools performance for a selection of the file types included in our benchmark - n/a indicates the tool doesn’t detect the given file type. |
Internally, Magika is used at scale to help improve Google users’ safety by routing Gmail, Drive, and Safe Browsing files to the proper security and content policy scanners. Looking at a weekly average of hundreds of billions of files reveals that Magika improves file type identification accuracy by 50% compared to our previous system that relied on handcrafted rules. In particular, this increase in accuracy allows us to scan 11% more files with our specialized malicious AI document scanners and reduce the number of unidentified files to 3%.
The upcoming integration of Magika with VirusTotal will complement the platform's existing Code Insight functionality, which employs Google's generative AI to analyze and detect malicious code. Magika will act as a pre-filter before files are analyzed by Code Insight, improving the platform’s efficiency and accuracy. This integration, due to VirusTotal’s collaborative nature, directly contributes to the global cybersecurity ecosystem, fostering a safer digital environment.
By open-sourcing Magika, we aim to help other software improve their file identification accuracy and offer researchers a reliable method for identifying file types at scale.
Magika code and model are freely available starting today in Github under the Apache2 License. Magika can also quickly be installed as a standalone utility and python library via the pypi package manager by simply typing pip install magika with no GPU required. We also have an experimental npm package if you would like to use the TFJS version.
To learn more about how to use it, please refer to Magika documentation site.
Magika would not have been possible without the help of many people including: Ange Albertini, Loua Farah, Francois Galilee, Giancarlo Metitieri, Luca Invernizzi, Young Maeng, Alex Petit-Bianco , David Tao, Kurt Thomas, Amanda Walker
By Elie Bursztein – Cybersecurity AI Technical and Research Lead and Yanick Fratantonio – Cybersecurity Research Scientist
Today we are open-sourcing Magika, Google’s AI-powered file-type identification system, to help others accurately detect binary and textual file types. Under the hood, Magika employs a custom, highly optimized deep-learning model, enabling precise file identification within milliseconds, even when running on a CPU.
|
| Magika command line tool used to recognize a identify the type of a diverse set of files |
You can try the Magika web demo today, or install it as a Python library and standalone command line tool (output is showcased above) by using the standard command line pip install magika.
Since the early days of computing, accurately detecting file types has been crucial in determining how to process files. Linux comes equipped with libmagic and the file utility, which have served as the de facto standard for file type identification for over 50 years. Today web browsers, code editors, and countless other software rely on file-type detection to decide how to properly render a file. For example, modern code editors use file-type detection to choose which syntax coloring scheme to use as the developer starts typing in a new file.
Accurate file-type detection is a notoriously difficult problem because each file format has a different structure, or no structure at all. This is particularly challenging for textual formats and programming languages as they have very similar constructs. So far, libmagic and most other file-type-identification software have been relying on a handcrafted collection of heuristics and custom rules to detect each file format.
This manual approach is both time consuming and error prone as it is hard for humans to create generalized rules by hand. In particular for security applications, creating dependable detection is especially challenging as attackers are constantly attempting to confuse detection with adversarially-crafted payloads.
To address this issue and provide fast and accurate file-type detection we researched and developed Magika, a new AI powered file type detector. Under the hood, Magika uses a custom, highly optimized deep-learning model designed and trained using Keras that only weighs about 1MB. At inference time Magika uses Onnx as an inference engine to ensure files are identified in a matter of milliseconds, almost as fast as a non-AI tool even on CPU.
|
| Magika detection quality compared to other tools on our 1M files benchmark |
Performance wise, Magika, thanks to its AI model and large training dataset, is able to outperform other existing tools by about 20% when evaluated on a 1M files benchmark that encompasses over 100 file types. Breaking down by file type, as reported in the table below, we see even greater performance gains on textual files, including code files and configuration files that other tools can struggle with.
|
| Various file type identification tools performance for a selection of the file types included in our benchmark - n/a indicates the tool doesn’t detect the given file type. |
Internally, Magika is used at scale to help improve Google users’ safety by routing Gmail, Drive, and Safe Browsing files to the proper security and content policy scanners. Looking at a weekly average of hundreds of billions of files reveals that Magika improves file type identification accuracy by 50% compared to our previous system that relied on handcrafted rules. In particular, this increase in accuracy allows us to scan 11% more files with our specialized malicious AI document scanners and reduce the number of unidentified files to 3%.
The upcoming integration of Magika with VirusTotal will complement the platform's existing Code Insight functionality, which employs Google's generative AI to analyze and detect malicious code. Magika will act as a pre-filter before files are analyzed by Code Insight, improving the platform’s efficiency and accuracy. This integration, due to VirusTotal’s collaborative nature, directly contributes to the global cybersecurity ecosystem, fostering a safer digital environment.
By open-sourcing Magika, we aim to help other software improve their file identification accuracy and offer researchers a reliable method for identifying file types at scale.
Magika code and model are freely available starting today in Github under the Apache2 License. Magika can also quickly be installed as a standalone utility and python library via the pypi package manager by simply typing pip install magika with no GPU required. We also have an experimental npm package if you would like to use the TFJS version.
To learn more about how to use it, please refer to Magika documentation site.
Magika would not have been possible without the help of many people including: Ange Albertini, Loua Farah, Francois Galilee, Giancarlo Metitieri, Luca Invernizzi, Young Maeng, Alex Petit-Bianco , David Tao, Kurt Thomas, Amanda Walker
By Elie Bursztein – Cybersecurity AI Technical and Research Lead and Yanick Fratantonio – Cybersecurity Research Scientist
We’re launching the AI Cyber Defense Initiative to help transform cybersecurity and use AI to reverse the dynamic known as the “Defender’s Dilemma”
A few weeks ago, we announced a series of product updates to Project IDX to help streamline and simplify full-stack, multiplatform software development. This week, we’re excited to share how Project IDX uses Gemini models to provide you with AI features to further speed up and refine your end-to-end developer workflow.
Project IDX launched with support for AI-powered code completion, an assistive chatbot, and contextual code actions like "add comments" and “explain this code” to help you write high-quality code faster. Since launch, and thanks to your feedback, we’ve been working hard to add new AI functionality to help boost your productivity even more.
You can now get inline AI assistance inside any file by pressing Cmd/Ctrl + I. Simply describe the changes you want to make to your code and IDX inline AI assistance will provide real-time error correction, code suggestions, and auto-completion in your code.
We integrated these AI enhancements directly into Project IDX’s centralized workspace to equip you with the necessary tools and resources for full-stack app development where and when you need them. From setting up your workspace to testing your app, IDX AI assistance helps accelerate and improve your workflow, ensuring that your end-to-end development experience is faster, easier, and higher quality.
For example, let’s say you want to add an authenticated API endpoint to your server. You can tell IDX AI to write the code necessary to enable secure task management using Firebase Authentication and Cloud Firestore. Given an input prompt, IDX AI assistance can write the code to construct the route, determine which APIs to use to verify the token, and save the data to the database. Instead of writing boilerplate code, you can focus on higher-level design and problem solving.
 |
| Input prompt for reference: Create a POST endpoint named /tasks. Get the ID Token from a cookie named _session. Verify this token with the Firebase Admin SDK. Use the UID property to assign the item to the user. Then save a task item with a servertime stamp for createdAt to the Firestore database using the admin SDK. |
Then, let's say you want to clean up your code a bit to improve its quality, readability, and maintainability. IDX AI assistance can help you quickly and easily refactor your code, so you can get right into optimizing your work without the hassle of manual refactoring.
 |
| Input prompt for reference: Refactor to use Node’s promise API. |
And, as you wrap up your project, IDX AI can help you test and debug your code to make sure your application is running smoothly before deployment. Tell IDX AI assistance to write you a unit test for a function to ensure it’s working properly, saving you time and effort as you inspect the quality of your app.
 |
| Input prompt for reference: Create a unit test for this function |
We’re also simplifying the process of building with the Gemini API with Project IDX’s new Gemini API template. The Gemini API template uses the Gemini Pro model to embed AI-powered features into your applications without additional configuration on your end, so you can get started working with the Gemini API quickly and easily. There's even an option to use the Gemini API via the popular LangChain framework to simplify the process of building LLM-powered apps.
The Gemini API template is multimodal, meaning it can provide context-aware prompt output for a myriad of input modalities including images, text and, of course, code. This can help you add features like conversational interfaces, summarization of user reviews, translation, and automatic image caption creation.
To demonstrate its functionality, we pre-configured the Gemini API template with ‘Baking with the Gemini API’, a recipe builder application that, using the Gemini model’s multimodal capabilities, can reverse-engineer possible recipes for baked goods from just a picture.
 |
But this recipe builder is just one example of the Gemini API template in action – with support for different input modalities and context-aware output generation, you can use IDX’s Gemini API template to create a myriad of innovative and impactful applications that deliver AI-enhanced experiences to your users.
These updates are a continuation of our efforts to leverage Google’s AI innovations for Project IDX, so make sure to keep an eye out for more announcements to come, including the expansion of AI in IDX to more than 150 countries/regions in the coming weeks.
Thank you for your continued support and engagement – please keep the feedback coming by filing bugs and feature requests. For walkthroughs and more information on all the features mentioned above, check out our documentation. If you haven’t already, visit our website to sign up to try Project IDX and join us on our journey. Also, be sure to check out our new Project IDX Blog for the latest product announcements and updates from the team.
We can’t wait to see what you create with Project IDX!
Gemini 1.5 delivers dramatically enhanced performance, with a breakthrough in long-context understanding across modalities.
Posted by Ali Satter – AI Lead, Roman Nurik – Design Lead