Together AI
By Together AI
Together AI typical implementation process:
Sign up for a Together AI account and obtain an API key from the dashboard; this typically takes a few minutes.
Set the TOGETHER_API_KEY environment variable and install an SDK (e.g., pip install together for Python or the TypeScript/AI SDK provider) or plan to call the HTTP API directly; expect 10–30 minutes.
Run a first test request against a chat or completion model (for example, GPT‑OSS or Llama) using the quickstart sample code to validate connectivity and streaming; most teams do this in under an hour.
Integrate Together AI into your application stack (Next.js, backend service, LangChain, Appwrite, etc.) by wiring the API key, choosing models, and adding basic error handling and logging; simple prototypes usually come together within 1–3 days.
Together AI can be customized deeply across deployment, models, performance, and workflow to fit specific business needs. Key customization options include:
Deployment environment: Run Together AI in the public Together Cloud, in your own VPC on any major cloud (AWS, Azure, GCP, OCI), or fully on‑premises, with data kept inside your firewall and full control over ingress/egress.
Model choice and ownership: Use 200+ open models (Llama, Mixtral, Gemma, DeepSeek, etc.), bring your own proprietary models, or train from scratch while retaining full control over model weights and IP.
Fine‑tuning behavior: Fine‑tune open‑weight models on your domain data, use Direct Preference Optimization (DPO), message weighting, and continued training to shape tone, safety, and task performance over time.
Deployment topology: Choose serverless endpoints for bursty workloads or dedicated instances/GPU clusters for predictable SLAs, latency, and isolation, with advanced GPU orchestration tuned to your traffic patterns.
Private cloud performance tuning: In private/VPC deployments, co‑design hardware–software configurations, batch strategies, and scaling policies to maximize specific accelerators and optimize training/inference costs.
Multi‑model orchestration: Combine and route between multiple models (open, custom, third‑party) using Mixture of Agents patterns to handle complex workflows, ensemble strategies, or tiered quality/cost routing.
Continuous optimization: Enable auto fine‑tuning, adaptive speculators (e.g., ATLAS), and distillation to keep models improving based on new data and changing usage rather than remaining static.
Together AI offers layered documentation, community resources, and formal support tiers so new users can get from first call to production reliably.
Together AI uses layered technical, organizational, and deployment controls to protect data across its cloud and private deployments.
Together AI ships updates continuously, with visible changes landing multiple times per month across its APIs, SDKs, and model library.
The public changelog shows a steady cadence of releases, including major feature drops (e.g., new video generation API, 40+ new image/video models, expanded audio streaming), plus frequent incremental improvements such as batch API upgrades, fine-tuning platform enhancements, and new model variants. For example, in September 2025 alone, Together AI rolled out a massively scaled batch inference API (raising limits from 10M to 30B enqueued tokens), upgraded fine‑tuning context lengths and Hugging Face integrations, and added new Qwen3 models.
Together AI’s policy is that you control your content and can choose whether or not it is retained or reused, with explicit mechanisms for portability and a “Zero Data Retention” mode for maximum control.
Together AI’s terms distinguish between your content (“Your Content”) and platform-generated or aggregated technical metrics (“Usage Data”). You keep ownership of Your Content, while Together may generate and use de-identified Usage Data (logs, performance stats, model-usage metrics) to operate, secure, and improve the services, without using it in a way that identifies you.
You can opt into or out of using your prompts and outputs for model training and product improvement via privacy settings. When Zero Data Retention (ZDR) is enabled, prompts, inputs, and outputs are not stored or used for training or product improvements, and are kept only as long as needed to process the specific request. This gives you fine-grained control over whether your operational data contributes to model learning or remains strictly ephemeral.
The privacy policy explicitly supports data access and portability rights (for example, under CCPA), allowing you to request copies of the personal data Together holds and to exercise data portability up to twice in a 12‑month period. You may ask for details about categories of data collected, purposes, and third parties, and request specific data points, which you can then export into your own systems.
Together AI uses a relatively provider‑friendly Terms of Service for its cloud platform, with flexible usage-based access but less customer control over renewal/cancellation than some SaaS tools.
Together AI meets key security and privacy standards focused on protecting customer data and enabling use in regulated industries, particularly healthcare.
Together AI has achieved SOC 2 Type 2 compliance, meaning an independent auditor has validated the design and operating effectiveness of its security controls—covering access management, data encryption, incident response, and change management—over an extended review period. This certification indicates adherence to the AICPA Trust Services Criteria around security, availability, and confidentiality, giving enterprises an audit-backed assurance of the platform’s security posture.
