Together AI is a cloud service provider that offers a platform for developing and running artificial intelligence models, with a particular focus on large language models and generative AI. The company provides access to GPU clusters and a suite of software tools designed to streamline the process of training, fine-tuning, and deploying models. Its offerings include the RedPajama series of open-source models, which are intended to serve as transparent and accessible alternatives to proprietary systems.
The platform is built around the principles of open research and reproducibility, aiming to lower the barriers to entry for organizations and researchers working in the field of Machine learning. By combining raw computing power with a managed software stack, Together AI seeks to address the complexities of infrastructure management that often hinder AI development. The company's approach has positioned it as a notable player in the competitive landscape of AI cloud services, which includes major providers like Amazon Web Services, Microsoft Azure, and Google Cloud.
History and Founding
Together AI was founded in 2022 by a group of researchers and engineers with backgrounds in artificial intelligence and distributed systems. The founding team included individuals who had previously worked at prominent technology companies and research institutions, bringing experience in large-scale model training and deployment. The company emerged from a desire to create a more collaborative and open environment for AI development, contrasting with the increasingly closed nature of some leading AI labs.
The initial funding round was completed in 2022, with investments from venture capital firms and notable tech industry figures. This early support allowed the company to begin building its cloud infrastructure and developing its initial software tools. By early 2023, Together AI had launched its public platform, offering access to GPU clusters and a range of open-source models.
RedPajama Models
One of the most significant contributions from Together AI is the RedPajama project, an initiative to create fully open-source large language models. The project's goal is to provide models that are not only free to use but also transparent in their training data and methodology. This contrasts with many commercial models, where the training data and internal workings are often proprietary.
The RedPajama dataset was released in April 2023, consisting of over 1.2 trillion tokens of text data. This dataset was designed to replicate the composition of the data used to train OpenAI's GPT-3, drawing from sources such as books, Wikipedia, and web pages. The release of this dataset was a major step in democratizing access to high-quality training data for large models.
Following the dataset release, Together AI introduced the RedPajama-INCITE family of models. These models, available in sizes from 3 billion to 7 billion parameters, were trained on the RedPajama dataset. They were made available under an open license, allowing for both research and commercial use. The models demonstrated competitive performance on various benchmarks, showing that open-source models could rival proprietary ones in certain tasks.
Cloud Platform and Infrastructure
Together AI's core product is its cloud platform, which provides users with on-demand access to high-performance computing resources. The platform supports a variety of GPU types, including those from AMD and NVIDIA, allowing users to select hardware that best fits their needs and budget. This flexibility is a key selling point, as it enables cost-effective experimentation and scaling.
The platform includes a managed training service that simplifies the process of training large models. Users can specify their model architecture and training parameters, and the platform handles the underlying infrastructure, including distributed training across multiple GPUs. This abstraction reduces the technical burden on researchers and developers, allowing them to focus on model design rather than cluster management.
For inference, Together AI offers a deployment service that allows users to run their trained models in production with minimal setup. The service includes features such as automatic scaling and load balancing, ensuring that applications can handle varying levels of traffic. The platform also supports fine-tuning, enabling users to adapt pre-trained models to specific domains or tasks.
Software and Tools
Beyond raw infrastructure, Together AI provides a suite of software tools designed to enhance the AI development workflow. These tools include libraries for model parallelism, which enable efficient training of very large models that exceed the memory capacity of a single GPU. The company has contributed to open-source projects in this area, such as the PyTorch-based libraries for distributed training.
The platform also integrates with popular frameworks like Transformer (architecture) architectures and Deep learning libraries, ensuring compatibility with existing workflows. Users can leverage standard tools like Hugging Face's Transformers library to load and use models hosted on Together AI. This interoperability is crucial for adoption, as it allows teams to migrate their existing code with minimal changes.
Together AI has also developed its own inference engine, optimized for low latency and high throughput. This engine is designed to serve large models efficiently, reducing the cost per token for users. The company claims significant performance improvements over generic serving solutions, making it an attractive option for applications with real-time requirements.
Open Source Contributions
Together AI is a strong proponent of open-source AI, and its contributions extend beyond the RedPajama models. The company has released several software libraries and tools under open-source licenses, aiming to benefit the broader AI community. These include utilities for data processing, model evaluation, and training optimization.
The company also participates in collaborative research efforts, publishing papers and technical reports on topics related to large-scale model training. This research is often conducted in partnership with academic institutions, such as Stanford AI Lab and BAIR (Berkeley AI Research). By sharing its findings, Together AI contributes to the collective knowledge base of the field.
One notable contribution is the FlashAttention library, which was developed in collaboration with researchers from Stanford AI Lab. FlashAttention is an efficient attention mechanism that reduces memory usage and speeds up training for transformer models. This library has been widely adopted in the AI community, becoming a standard component in many training pipelines.
Competitive Landscape
Together AI operates in a highly competitive market, facing challenges from both established cloud providers and specialized AI startups. Major cloud platforms like Amazon Web Services, Microsoft Azure, and Google Cloud offer their own AI services, including managed machine learning platforms and pre-trained models. These providers have the advantage of vast infrastructure and existing customer bases.
Specialized competitors include Groq, SambaNova, and Graphcore, which focus on developing custom hardware and software for AI workloads. These companies often target high-performance inference and training, offering alternatives to traditional GPU-based systems. Together AI differentiates itself through its emphasis on open-source models and its user-friendly platform.
The company also faces competition from other open-source model providers, such as AI21 Labs and Inflection AI. However, Together AI's unique combination of cloud infrastructure and model development sets it apart. The company aims to be a one-stop shop for organizations looking to build and deploy custom AI solutions.
Applications and Use Cases
Together AI's platform is used across a wide range of industries and applications. In the healthcare sector, organizations use the platform to develop models for medical text analysis and diagnostic support. The ability to fine-tune models on proprietary data is particularly valuable in this domain, where data privacy is paramount.
In the financial industry, Together AI helps companies build models for fraud detection, risk assessment, and algorithmic trading. The platform's scalability allows these organizations to process large volumes of data and deploy models in real-time. The open-source nature of the models also enables greater transparency and auditability, which is important for regulatory compliance.
For startups and research institutions, Together AI provides an accessible entry point into large-scale AI development. The pay-as-you-go pricing model allows teams to experiment without significant upfront investment. This democratization of access is a core part of the company's mission, enabling a broader range of voices to contribute to AI innovation.
Future Directions
Looking ahead, Together AI continues to expand its platform and model offerings. The company is investing in research to improve the efficiency and capabilities of large language models, with a focus on reducing the computational cost of training and inference. This includes exploring techniques like model pruning and quantization, which can make models smaller and faster without significant loss in performance.
The company is also working on enhancing its support for multimodal models, which can process both text and images. This is an area of growing interest in the AI community, with applications in fields like computer vision and robotics. By expanding its capabilities, Together AI aims to remain at the forefront of the generative AI revolution.
As of 2024, Together AI has established itself as a key player in the AI cloud market, with a growing customer base and a strong reputation for open-source contributions. The company's continued success will depend on its ability to innovate and adapt to the rapidly evolving landscape of artificial intelligence.