January 2025 Summaries

DeepSeek-R1 is an open-weight competitor to proprietary reasoning models like OpenAI's o1, delivering powerful reasoning at a fraction of the cost. Together AI offers both the full R1 and distilled models with opt-out privacy controls and serverless pay-per-token pricing, allowing developers to experiment freely without costly GPU deployments. The main DeepSeek-R1 model rivals OpenAI's o1 in reasoning tasks while running 9x cheaper on high-performance serverless infrastructure at a competitive rate of $7 per 1 million tokens. Serverless deployment for R1 matters due to its large size and the advantages it offers, including pay-per-token pricing, high-performance infrastructure, and full flexibility to scale deployments as needed. The distilled models are not the main DeepSeek-R1 model but rather fine-tuned variants that have been trained with reasoning examples generated by DeepSeek-R1. Together AI prioritizes security with opt-out privacy controls and hosts all R1 models in its own data centers ensuring sensitive information remains secure. A free endpoint for DeepSeek-R1 Llama 70B distilled is also available, offering easy access to powerful reasoning models while notes that the free model endpoint has reduced rate limits and performance compared to paid Turbo endpoints.

Mistral Small 3, a new open-source model, is now available on Together AI, offering performance comparable to larger models like Llama 3.3 70B at just 24B parameters. This smaller model delivers great results for general-purpose conversational assistance and low-latency applications such as robotics. Developers can integrate Mistral Small 3 into their existing applications using Together AI's OpenAI-compatible APIs. The model is available on Together Serverless with pay-per-token endpoints or via scalable dedicated GPU endpoints, making it suitable for production workloads. With its smaller size and comparable performance, Mistral Small 3 enables developers to run high-quality AI with faster response times and lower costs, driving innovation in the industry.

Flux LoRAs are a powerful way to customize AI image generation for specific styles by specifying one or more Flux LoRA adapters during image generation. The Together API supports using Flux LoRA models, and popular LoRAs such as Flux Sketch LoRA, Flux Logo Design LoRA, Flux Colored Sketch LoRA, Flux Outfit Generator LoRA, and Flux Anime Sketch LoRA can be used to generate images with specific styles. These models can be trained on a small number of images and provide a convenient way to reproduce specific visual styles. Users can also use their own custom LoRA models by training them themselves and uploading them to HuggingFace. The Together Platform provides a serverless platform for running inference against any Flux LoRA, with pay-per-token pricing and no need for additional infrastructure. To get started with Flux LoRAs on the Together AI platform, users can sign up, upload existing LoRA adapters, follow along with a Python notebook, or read the docs to start using the Together API.

Today, Together AI is announcing the launch of its new audio API powered by Cartesia Sonic, a state-of-the-art low-latency and ultra-realistic voice model. This partnership allows developers to access the Sonic model directly via the Together API, expanding the platform's capabilities to build multi-modal applications with chat, image, audio, code, and embeddings through a single platform. The addition of audio capabilities marks a significant milestone in enabling developers to orchestrate multiple AI models working together for voice applications, such as conversational experiences, customer support, content creation, and digital twins. Cartesia Sonic consistently outperforms other voice models with ultra-low latency, superior content processing, and leading research breakthroughs, making it an attractive choice for developers building voice AI applications. With the Together Platform, developers can now build enterprise-ready voice applications on HIPAA and SOC2 compliant systems, get started with voice applications using cookbooks, and orchestrate multiple AI models for seamless interactions.