Together AI

Together AI is an AI cloud platform that provides fast, affordable access to open-source large language models for both inference (running AI to generate responses) and fine-tuning (customising models on your own data). It gives developers and enterprises a simple API to access over 100 open-source models — including Llama 3, Mistral, Qwen, DeepSeek, and Code Llama — without managing GPU infrastructure. Together AI is known for industry-leading inference speeds and prices significantly lower than closed API providers, making it the go-to platform for teams who want open-source AI model flexibility without infrastructure complexity.

OpenAI provides access to its own proprietary, closed-source models (GPT-4, GPT-4o, o1) that you cannot inspect, modify, or run yourself. Together AI provides access to open-source models (Llama, Mistral, Qwen, etc.) that are publicly available and can be run anywhere. Key differences: Cost — Together AI's models are often 10-25× cheaper per token than comparable OpenAI models. Control — open-source models can be fine-tuned on your data and customised in ways closed models cannot. Privacy — with dedicated deployments, your data does not flow through a shared model. Vendor lock-in — open-source models are not controlled by any single company. Importantly, Together AI's API is OpenAI-compatible, meaning switching from OpenAI to Together AI requires changing only one line of code in most applications.

Together AI offers $25 in free credits for new accounts — enough to run millions of tokens for testing and prototyping. After the free credits are used, pricing is pay-per-token with no minimum monthly commitment. Prices vary by model — smaller models (like Llama 3 8B) cost significantly less than larger models (like Llama 3 70B). Some models are offered for free on Together AI's platform. Enterprise customers with high-volume needs can negotiate custom pricing and dedicated deployments. Visit together.ai/pricing for current rates, which are updated as infrastructure costs and competition evolve.

Together AI hosts over 100 open-source models including: Llama 3 (Meta) — 8B and 70B instruction-tuned variants, the current flagship open-source LLM family. Mistral — including Mistral 7B, Mixtral 8x7B, and Mistral Large. Qwen — Alibaba's multilingual model family with strong non-English performance. DeepSeek — including DeepSeek Coder for programming tasks. Code Llama — Meta's coding-specialised Llama variant. Falcon — from the Technology Innovation Institute. DBRX — Databricks' high-performance model. Embedding models (for semantic search and RAG applications). The catalogue is updated regularly — new models are typically available within hours of public release. Check together.ai/models for the current full catalogue.

Together AI was co-founded by Vipul Ved Prakash (CEO) and Ce Zhang (CTO), along with other co-founders. Vipul is a serial entrepreneur with experience in internet infrastructure and distributed systems — previously co-founded Topsy Labs (acquired by Apple in 2013) and has deep expertise in building systems that process data at massive scale. Ce Zhang is a professor of Computer Science at ETH Zurich and the University of Chicago, with internationally recognised research on scalable machine learning systems and data management for AI. His academic expertise directly informs Together AI's engineering approach to high-performance AI inference. You can learn more about both founders in the Founders section above.

Yes — fine-tuning is one of Together AI's core capabilities. You can fine-tune supported open-source models (including Llama and Mistral variants) on your own training data through Together AI's API or web interface. The process involves uploading your training data in JSONL format (prompt-completion pairs), configuring training parameters (learning rate, epochs, batch size), and launching the training job. Together AI handles all the GPU compute and infrastructure. The resulting fine-tuned model is hosted on Together AI and accessible via your private API key. Fine-tuning is typically charged per GPU hour used for training plus ongoing inference costs for the hosted model.

Together AI consistently achieves some of the highest tokens-per-second rates in independent benchmarks. For Llama 3 70B, Together AI has achieved over 100 tokens per second in benchmark tests — significantly faster than most alternative providers. Speed matters for user-facing applications where AI response latency directly affects user experience, and for batch processing applications where higher throughput means lower total processing time and cost. Together AI's CTO Ce Zhang's background in AI systems research — specifically in making ML inference more efficient — is directly responsible for this performance advantage. Actual performance varies by model size, traffic levels, and whether you are using shared or dedicated infrastructure.

Yes — Together AI has an enterprise offering designed for large organisations with production AI workloads. Enterprise features include: Dedicated deployments (GPU infrastructure reserved exclusively for your organisation, providing guaranteed performance SLAs and data isolation). SOC 2 Type II compliance (certified security controls). HIPAA-compatible configurations for healthcare applications. Enterprise SLAs with uptime guarantees and dedicated support. Custom pricing for high-volume customers. SSO and access management integrations. Many Fortune 500 companies and large enterprises use Together AI for production AI workloads — particularly those that need open-source model flexibility, data privacy, or regulatory compliance that is difficult to achieve with closed API providers.

Inference is using an existing AI model to generate responses — you send a text prompt, the model processes it, and returns a text output. This is what happens every time you use ChatGPT or any AI chatbot: the model is doing inference. Fine-tuning is a process of training an existing model further on a new dataset to specialise its behaviour for a specific task, domain, or style. For example: starting with a general-purpose Llama 3 model and fine-tuning it on thousands of examples of your company's customer service conversations would produce a model that is better at handling your specific customer queries than the generic model. Together AI provides infrastructure for both: inference (running models at scale through the API) and fine-tuning (training custom models on your data and hosting them).

Open-source AI models offer several important advantages over closed models. Cost — as noted, 10-25× cheaper per token in many cases. Customisation — you can fine-tune open-source models on your data in ways that are not possible with closed models. Data privacy — with dedicated deployments, your data stays in your own cloud instance rather than passing through a shared service. No vendor lock-in — you are not dependent on a single company's pricing, availability, or API changes. Transparency — open-source model weights are publicly available for inspection by researchers and security teams. Compliance — for regulated industries, self-hostable open-source models offer clearer data governance paths than closed APIs. The trade-off has historically been that closed frontier models (GPT-4o, Claude 3.5) slightly outperform open-source alternatives on complex reasoning tasks — but this gap has narrowed significantly with each generation of open-source models, and for most production use cases, open-source models on Together AI perform comparably at a fraction of the cost.