Together AI

Together AI is a cloud platform specialized in running and fine-tuning open source large language models with high performance and competitive pricing. By focusing exclusively on open source AI inference and training, Together AI provides developers and enterprises with fast, cost-effective access to the latest open source foundation models without the complexity of managing GPU infrastructure.

The Together AI inference platform offers one of the most comprehensive catalogs of hosted open source LLMs available. The catalog includes Meta Llama 3, Mistral, Mixtral, Qwen, DeepSeek, Gemma, Phi, Code Llama, and many other models across various sizes and specializations. Models are available for text generation, code generation, function calling, and vision tasks, covering the broad spectrum of LLM application requirements.

Speed is a defining characteristic of Together AI's infrastructure. The platform uses custom inference optimizations, speculative decoding, and specialized hardware configurations to achieve very high token generation speeds, often significantly faster than comparable infrastructure from general-purpose cloud providers. For applications where LLM response latency is a critical user experience factor, Together AI's performance advantage can be meaningful.

The Together AI API follows the OpenAI-compatible format, meaning that applications already using the OpenAI API can switch to Together AI by changing the base URL and API key, with no other code changes required. This compatibility dramatically reduces the effort required to integrate Together AI into existing applications or compare its performance against other providers.

Fine-tuning on Together AI enables organizations to adapt foundation models to their specific use cases, domain vocabulary, and desired response styles. The platform supports supervised fine-tuning using datasets uploaded in JSONL format, with automatic model evaluation and deployment upon completion. Fine-tuned models are private to the organization and accessible through the standard API.

Together AI also provides training capabilities for organizations that want to pre-train custom models or perform continued pre-training on proprietary data. The training platform handles distributed training across multiple GPUs and nodes, with checkpointing, failure recovery, and monitoring built in.

The pricing model is usage-based, with rates per million input and output tokens that are typically lower than comparable proprietary API providers. Different model sizes and architectures are priced differently, enabling teams to optimize for the right balance of cost and capability for each use case.

Together AI is well suited for AI startups, research teams, and enterprises that want the flexibility of open source models, the compliance benefits of controlling model selection, and the cost advantages of competitive pricing, all without the operational burden of managing GPU infrastructure.