DeepInfra

Run open models on DeepInfra without hosting GPUs. Agents chat with Llama, Qwen, DeepSeek, and Mistral through the OpenAI-compatible API, create embeddings for RAG, generate images with FLUX or SDXL, and reach any hosted model via the native inference endpoint.

6 actions API key available

DeepInfra website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

The morning's inbound messages queue up from the shared [Gmail](/integrations/gmail-service) inbox

Simple Chat classifies and summarizes each message on an open model chosen from the live catalog

Create Embeddings indexes each message into [Pinecone](/integrations/pinecone) for the searchable history

The agent scores classification confidence and sets the uncertain items aside

An operator reviews the low-confidence classifications, and the corrections feed back into the prompt's examples

Summaries file to [Notion](/integrations/notion) with their labels attached

The triage digest posts to the operations channel in [Slack](/integrations/slack)

What This Integration Enables

DeepInfra's case is blunt: most workflow AI is routine work, classification, extraction, summarization, and routine work should run on open models at open-model prices, hosted by someone else. The connector exposes both of DeepInfra's surfaces over one catalog: the OpenAI-compatible API for chat and embeddings, where existing patterns port by changing the model id, and the native inference endpoint that reaches everything else the platform hosts, from zero-shot classifiers to speech and vision models. Service tiers turn urgency into a dial rather than an architecture. What stays human is the judgment call that matters: whether an open model's output quality earns the [human-in-the-loop](/concepts/human-in-the-loop) sign-off to serve a workload that used to run on a proprietary one. - Run chat, summarization, and extraction on Llama, Qwen, DeepSeek, Mistral, and Gemma without hosting anything - Build RAG on open embedding models with vectors returned in input order - Generate images with FLUX and SDXL, normalized to stable FlowRunner file URLs - Reach classifiers, speech, and vision models through the native inference passthrough

Without FlowRunner

Open models mean GPUs Running Llama or Qwen yourself is an infrastructure project before it is anything useful

Every call pays frontier rates Simple classification jobs run through the priciest models because that is what is wired up

Odd modalities need odd vendors A classifier here, a speech model there, each with its own integration

With FlowRunner

Open weights, serverless Per-token billing on hosted open models, with nothing to provision or patch

The tier matches the job Priority buys latency for interactive steps, Flex takes a discount on batch work

One catalog, every modality Chat, embeddings, images, classifiers, and speech through a single key

Use Case Scenarios

A RAG index priced for completeness

Embedding cost decides how much of your corpus actually gets indexed. The agent runs Create Embeddings on an open embedding model over incoming documents and upserts the vectors to [Pinecone](/integrations/pinecone), with the model id chosen from List Models by price and context length. Because the per-token cost sits well below proprietary embedding APIs, the index covers everything, including the archives that would otherwise be judged not worth embedding. Coverage decisions stop being budget decisions, and retrieval quality shows it.

Overnight batches on the discount tier

The nightly summarization run has no user waiting on it, so the agent submits it through Chat Completion on the Flex service tier, trading best-effort scheduling for a discounted rate. Interactive steps in daytime flows run Priority when the latency is genuinely worth paying the surcharge for. Same models, same prompts, different economics per job, and the tier is a parameter the flow sets per call rather than a commitment anyone renegotiates. The batch that finishes at 4 a.m. instead of 2 a.m. cost less, and nobody was waiting either way.

One key from artwork to transcripts

Campaign artwork renders through Text to Image on FLUX, with results normalized into stable file URLs and posted to [Slack](/integrations/slack) for review. When a workflow needs something no chat endpoint offers, a zero-shot classifier for routing, or speech transcription on an audio file, Run Inference passes the model-shaped request straight through and returns the raw result with its inference status and cost. The escape hatch means the odd jobs stop requiring odd vendors, odd contracts, and odd invoices, and the whole workflow reads from one place.

Human-in-Loop Highlight

The expensive mistake on DeepInfra is not a button, it is a migration. Moving a customer-facing workload from a proprietary model to an open one through Chat Completion is a one-field change that can cut unit cost dramatically, and that ease is precisely the trap: quality regressions in tone, edge-case handling, and instruction-following do not show up in a smoke test. So cutovers run as evaluations, not swaps. The agent runs the workload's real prompt set through both models, lays the outputs side by side with the cost delta, and the owner reads before anything serves customers. The same discipline prices urgency honestly: Priority's surcharge on a high-volume flow is a standing spend increase, and standing spend gets a named approver, not a default.

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

6 actions

Chat

Chat Completion Full chat completion against any hosted open text model through the OpenAI-compatible endpoint: multi-turn messages, sampling controls, stop sequences, structured outputs, tool calling, reasoning effort for reasoning models, and DeepInfra service tiers, Priority for lower latency at a surcharge and Flex for a discount with best-effort scheduling. Complete responses, never streamed, with most models capping generation at 16,384 output tokens.
Simple Chat One prompt in, the assistant's plain-text answer out, with optional system prompt and temperature and length controls. The right size for classification and extraction steps.

Embeddings

Create Embeddings Generates vectors for one or more texts on hosted open embedding models, one vector per input in input order, with model used and token usage returned. The RAG and dedup primitive.

Images

Text to Image Generates images from a prompt on FLUX, SDXL, and other hosted image models, with per-model knobs like dimensions, steps, seed, and, on SDXL, negative prompt and guidance scale. Both of DeepInfra's return forms are normalized and saved to FlowRunner file storage as stable URLs.

Inference

Run Inference Calls the native inference endpoint for any hosted model, passing the request body through unchanged, so the model defines the shape. The route to classifiers, token classifiers, fill-mask, zero-shot, detection, and speech models, with the raw response returned including runtime and cost.

Models

List Models Lists the hosted catalog with owner, description, context length, output limits, pricing, and category tags, narrowed by search, category, and limit. The discovery step, and deliberately so: model ids churn as models are added and retired, so flows pick from the live catalog instead of trusting a stale default.

Frequently Asked Questions

What can FlowRunner do with DeepInfra?

FlowRunner agents can run Chat Completion, Simple Chat, and Create Embeddings in DeepInfra, plus 3 more actions.

Does connecting DeepInfra to FlowRunner require OAuth?

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

Can DeepInfra trigger a FlowRunner workflow automatically?

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

Start building with DeepInfra

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

Start Building Free Talk to us