Together AI

Run open models on Together AI from your flows: chat and text completions across Llama, DeepSeek, Qwen and hundreds more, plus embeddings, document reranking, FLUX image generation, speech synthesis, and audio transcription and translation through one OpenAI-compatible API.

10 actions API key available

Together AI website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

A new support question arrives from the helpdesk

Create Embeddings converts the question into a vector for the knowledge base search

Rerank Documents reorders the vector store's candidate passages best-first, with a relevance score per document

The agent compares the top relevance score against the confidence floor

Chat Completion drafts a grounded answer with a json_schema response format, citing the passages it used

Questions that score below the floor route to a support engineer instead of a guess, and the gap is filed for the docs team

The drafted answer posts to the helpdesk thread with its sources and scores attached for the agent handling the ticket

What This Integration Enables

The model layer is commoditizing, and Together AI is what that looks like as infrastructure: hundreds of open models, Llama, DeepSeek, Qwen, Mixtral, GPT-OSS, FLUX, Whisper, behind one OpenAI-compatible API, with the license, context window, and per-token price published for every one. For a team building automation, that changes the shape of the decision. The pipeline is the asset; the model is a component you shop for, benchmark, and swap when the leaderboard moves. FlowRunner's BYOK approach is built for exactly this posture: your Together AI key, your model choices, orchestrated inside flows that outlive any single model generation. - Run one-shot prompts with Ask, full multi-turn conversations with Chat Completion, and raw continuations with Text Completion for base models and infill work - Build the complete RAG loop natively: Create Embeddings for recall, Rerank Documents for precision, structured outputs for the answer - Generate FLUX images and multi-provider speech, saved to FlowRunner file storage with permanent URLs - Transcribe audio with speaker diarization and word-level timestamps, or translate spoken audio straight into English - Interrogate the model catalog itself, filtering by type and checking pricing before a single token is spent

Without FlowRunner

One-model monogamy The whole stack is welded to a single provider's model, pricing, and roadmap

RAG that recalls but cannot rank Vector search returns twenty near-matches and the answer gets built on the wrong one

Model choice by press release Which model runs in production was decided by whatever was loudest that month

With FlowRunner

A model per task Chat on Llama or DeepSeek, embeddings on BGE, precision on a reranker, images on FLUX, audio on Whisper, one API

Recall, then precision Embeddings cast the wide net and Rerank Documents sharpens the final ordering before anything is generated

Choice with the price sheet open List Models returns context windows and per-million-token pricing, so the pick is a calculation

Use Case Scenarios

The recordings folder that writes its own minutes

When a new recording lands in [Dropbox](/integrations/dropbox-service), the flow passes its link to Transcribe Audio with diarization on, so the transcript says who spoke, not just what was said. Ask condenses the timestamped transcript into decisions, owners, and deadlines, and the summary files itself into [Notion](/integrations/notion) with the full transcript attached. The meeting's obligations stop depending on whoever took notes.

Support search with a precision stage

The naive RAG flow embeds a question, takes the top vector hit, and hopes. This one adds the stage that makes it production-grade: candidates from [Qdrant](/integrations/qdrant) pass through Rerank Documents, where a purpose-built rerank model scores true relevance, and only the strongest passages reach Chat Completion. The answer arrives with citations, and the relevance scores give the flow an honest way to know when it should not answer at all.

Campaign assets on open weights

A new record in the campaign table in [Airtable](/integrations/airtable) kicks off asset production: Generate Image renders hero art with FLUX, seeded for reproducibility, and Generate Speech produces the voiceover from the approved script. Both land in FlowRunner file storage, and both post to [Slack](/integrations/slack) where a marketer approves before anything reaches a customer-facing channel. The generation is cheap and fast; the judgment stays human.

Human-in-Loop Highlight

The commitment hiding in this catalog is Create Embeddings at corpus scale. Embedding one question costs fractions of a cent; re-embedding the entire knowledge base is a real bill, and it is never just a bill, because vectors from different embedding models are incompatible. Choosing to move from one embedding model to another means regenerating every vector in the index and cutting over the search that production flows depend on. So when an agent's retrieval-quality report suggests a better embedding model exists, it does not migrate. It presents the case: the candidate model's scores on a sample, the corpus size, the estimated token spend from List Models pricing, and the cutover plan. A person approves the migration window. The same discipline applies at smaller scale to bulk Generate Image runs, where a mistyped batch multiplies FLUX spend silently. Inference is cheap enough to waste only until it is not; the budget-shaped decisions stay with the budget's owner.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

10 actions

Chat

Ask Sends a single prompt to any open model on Together AI and returns just the answer as plain text, with an optional system prompt and a JSON mode for parseable replies. The convenience action for one-shot generation steps.
Chat Completion Generates a response for a full chat conversation on any hosted model, with the complete messages array, the full sampling surface, structured outputs via text, json_object, or json_schema response formats, tool calling, an optional moderation safety model, and over-length context handling. Returns the raw response with tool calls, finish reason, and token usage.
Text Completion Generates a raw continuation with no chat template applied, suited to base models, few-shot prompting, code infilling, and fill-in-the-middle work. Supports multiple completions per request and prompt echo.

Embeddings

Create Embeddings Generates embedding vectors for one or more texts using any hosted embedding model (BGE, GTE, M2-BERT, multilingual E5, and more), one vector per input in order. The recall half of every RAG pipeline, and the action this page's human gate watches at corpus scale.
Rerank Documents Reorders documents by semantic relevance to a query using a hosted rerank model, returning each document's original index and relevance score sorted best-first. The precision step after vector search.

Images

Generate Image Generates images from a text prompt with the FLUX.1 family, with explicit dimensions, step count, guidance scale, negative prompts, deterministic seeds, and an optional input image for image-to-image work. Saves to FlowRunner file storage by default, with provider URL and base64 delivery as alternatives.

Audio

Generate Speech Converts text to speech using hosted TTS models (Cartesia Sonic, Kokoro, Orpheus, Rime, MiniMax, Deepgram Aura, and more), with the voice chosen from each model's live published catalog. Output in MP3, WAV, or raw PCM, saved to FlowRunner file storage.
Transcribe Audio Transcribes an audio file in its original language using Whisper Large v3, from a FlowRunner file URL or any public URL, with language hints, a decoding bias prompt, verbose segment and word timestamps, and speaker diarization with optional speaker-count bounds.
Translate Audio Translates spoken audio in any supported language into English text using Whisper Large v3, with a source language hint and timestamped verbose output.

Models

List Models Lists the models Together AI serves with full metadata: ID, display name, hosting organization, type, context window, license, and per-million-token pricing, with type filters and free-text search. The action that turns model selection into an informed decision.

Frequently Asked Questions

What can FlowRunner do with Together AI?

FlowRunner agents can run Chat Completion, Ask, and Text Completion in Together AI, plus 7 more actions.

Does connecting Together AI to FlowRunner require OAuth?

No. Together AI connects to FlowRunner with an API key, no OAuth flow required.

Can Together AI trigger a FlowRunner workflow automatically?

Together AI doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with Together AI

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us