Skip to main content
Together AI hosts one of the widest catalogs of open-weight models behind a single OpenAI-compatible API. Its serverless tier is per-token with no provisioning, replicas or minimum spend, so you can try several open models on the same agent in an afternoon and keep the one that sounds best on calls. Two things are specific to Together and catch people out:
  • Model IDs are namespaced as <organisation>/<model_name>, for example MiniMaxAI/MiniMax-M3, not OpenAI-style flat names.
  • Serverless performance is best-effort. For guaranteed throughput, Together offers provisioned throughput and dedicated endpoints, whose model catalog is different from the serverless one.

​
Set up Together AI

1

Create an account and an API key

Register at api.together.ai, then open your project’s API keys settings page and copy a key.
2

Choose a model from the serverless catalog

Open Together’s serverless models list and copy the model ID from the chat models section. Ignore image, audio and embedding models; Bolna needs a chat model. Together’s recommended models page is a good shortlist.
3

Choose Together AI in Bolna

Go to Settings → Integrations → LLM → Custom LLM → Add model, or choose + Add custom LLM from the LLM Provider dropdown on an agent. Select Together AI.
Bolna Add custom model dialog with Together AI selected and api.together.ai/v1 shown as the endpoint placeholder

Together AI selected in the Add custom model dialog

4

Register the model

  • Model name: the full Together model ID, for example MiniMaxAI/MiniMax-M3
  • Endpoint URL: https://api.together.ai/v1
  • API key: your Together API key
  • Family: the model’s family, for example minimax, llama or qwen
Add custom model dialog with Together AI selected, model MiniMaxAI/MiniMax-M3, endpoint api.together.ai/v1, masked key and family minimax

Together AI with a namespaced model ID

5

Select it on the agent

On the Intelligence tab pick custom and the Together model, save, then call the agent.

​
Comparing several open models

Together makes A/B testing cheap. Register two or three models, for example a small fast model and a larger one, each as its own custom model in Bolna. Clone the agent, point each copy at a different model, and compare call recordings for reply latency, accuracy and how natural the speech sounds. Keep the winner and delete the rest from Settings → Integrations.

​
Rate limits and throughput

Most serverless traffic never hits a limit, but Together may rate-limit large bursts or apply custom limits to models in especially high demand. For a large outbound batch:
  • Spread the batch over time instead of launching every call at once.
  • Prefer a model that isn’t in peak demand.
  • For committed capacity, use Together’s provisioned throughput or a dedicated endpoint. Check the dedicated catalog first, as it doesn’t list the same models as serverless.

​
Troubleshooting

Include the organisation prefix and match the case exactly. MiniMax-M3 alone won’t resolve; MiniMaxAI/MiniMax-M3 will.
Serverless capacity is best-effort. Pace the batch, or move to provisioned throughput.
Keys are scoped to a Together project. Make sure the key belongs to the project with billing set up.