pAIring connectors

A connector lets pAIring send chat requests to models that run in the cloud, instead of a local llama.cpp or Matricxon. It is useful when you need a large model your own hardware cannot hold. All but xAI, Azure, Vertex AI, Bedrock and RunPod can be tried for free, with a daily allowance, free models, free credits or a trial key, before you pay for anything. Switching between them, or back to a local engine, is one setting.

Compare the connectors

Connector	What you enter	Models	Free use	Knowledge base	Good to know
Cloudflare Workers AI	Account ID, API token	Dozens: Gemma, Llama, Qwen, Mistral, GPT-OSS	10,000 neurons a day	Yes	Usage is shown in pAIring
DigitalOcean	Model access key	Dozens, including OpenAI and Anthropic models	Free tokens to start	Yes	Some listed models need a paid plan
OpenRouter	API key	Hundreds from many providers	Free models, with a daily request limit	Yes (paid embedding models)	Free models may log your prompts
Google Gemini	API key	Gemini and Gemma	Free tier	Yes	Free-tier prompts may be used by Google; paid tier only in the EEA, UK and Switzerland
Groq	API key	Fast open models: Llama, Qwen, GPT-OSS	Free tier, limits per model	No	Very fast replies
xAI Grok	API key	The Grok models, from fast to large reasoning ones	No, billed per use	No	Not the same company as Groq
NVIDIA NIM	API key	100+: Llama, Nemotron, Mistral, DeepSeek, Qwen	Free developer credits	Yes	Credits are meant for trying, not production
Cohere	API key	Command chat and Embed models	Trial key, about 1,000 calls a month	Yes	Trial keys are meant for trying, not production
Azure OpenAI	Endpoint, API key, deployment names	Your own deployments: GPT, DeepSeek, Grok, Llama and more	No, billed to your Azure subscription	Yes	Data stays in your Azure resource and region
Google Vertex AI	Service-account key, models	Gemini and Model Garden partner models	No, billed to your Google Cloud project	No	Your project’s region and data controls apply
Amazon Bedrock	Region, Bedrock API key or IAM access key, models	GPT-OSS, Claude, Llama, Mistral, DeepSeek, Qwen and more	No, billed to your AWS account	No	Use models whose card lists Chat Completions
RunPod Serverless	Endpoint ID, API key, model	The model your own endpoint serves	No, billed by use	No	You deploy a vLLM endpoint first

Which generation settings each connector passes on (the rest are left at the provider’s defaults):

Connector	Temperature, Top-P	Top-K	Repeat penalty	Reply length	Seed
Cloudflare Workers AI	Yes	Yes, up to 50	Yes	Yes	Yes
DigitalOcean	Yes	No	No	Yes	Yes
OpenRouter	Yes	Yes	Yes	Yes	Yes
Google Gemini	Yes	No	No	No	No
Groq	Yes	No	No	No	Yes
xAI Grok	Yes	No	No	No	Yes
NVIDIA NIM	Yes	No	No	Yes, at least 1,024 tokens	Yes
Cohere	Yes	No	No	No	Yes
Azure OpenAI	Yes, except on reasoning models	No	No	No	Yes, except on reasoning models
Google Vertex AI	Yes	No	No	No	Yes
Amazon Bedrock	Yes, where the model allows	No	No	No	Yes, where the model allows
RunPod Serverless	Yes	Yes	Yes	Yes	No

A reply-length limit is left out for some providers on purpose: their reasoning models think before they answer, and a short limit, such as a chat title’s, would leave the reply empty.

Admins open Connectors from the sidebar. Each connector is a card. Click it to open its settings.

Set one up

1. Open the connector’s card and fill in its fields.
2. Press Save.
3. Press Test connection. It must pass before the connector can be used. Run it again after any change.
4. Switch the connector on.
5. Open Admin settings → Inference engines and pick it as the active engine.

Secret fields, such as API keys and tokens, are stored encrypted and are not shown again.

Cloudflare Workers AI

Uses the models Cloudflare hosts on its own GPU network, through its OpenAI-compatible API. There is nothing to deploy.

Field	What to enter
Account ID	Your Cloudflare account ID. It is shown on the Workers AI page and on any domain’s Overview page
API token	A custom token from My Profile → API Tokens, for your account, with Workers AI: Read. Add Account Analytics: Read to see usage in pAIring

What works

• Many models. Chat models such as Gemma 4, Llama 4, Qwen, Mistral, DeepSeek and GPT-OSS. They appear in the model list, and each chat can use a different one.
• Streaming chat, with your generation settings. Temperature, Top-P, Top-K (up to 50), repeat penalty, reply length and seed are passed on.
• Knowledge base. Cloudflare’s embedding models work for document search. See Knowledge base below.
• Chat titles and other short internal requests.

Knowledge base (RAG)

Cloudflare can also turn your documents into searchable data, so chats can answer from them.

1. Open Admin settings → System → Default models and pick a Cloudflare embedding model as the Default embedding model.
2. Make sure RAG (Knowledge base) is switched on under Admin settings → System.
3. Add documents on the Knowledge tab, as with any engine.

Embedding model	Good for
@cf/baai/bge-m3	A good default: many languages, including Hebrew, and long passages
@cf/baai/bge-small-en-v1.5	The lightest and cheapest, English only
@cf/qwen/qwen3-embedding-0.6b	Many languages, small

Embedding uses very little of your daily allowance compared with chat.

Switching embedding models

Documents are searched only with the model they were embedded with. pAIring remembers that model for each document. After you switch, documents from the old model are left out of searches until they are embedded again, which the next sync does automatically. To do it at once, press Re-embed all documents under Admin settings → System → RAG.

Summary in Inference engines

When Cloudflare is the active engine, Admin settings → Inference engines shows a summary instead of a server form:

• Status: enabled, connection test passed, ready.
• Account: the account ID, and whether the token is active. The token itself is never shown.
• Models: how many chat and embedding models your account can use.
• Usage: requests, neurons and tokens over the last 7 days and today, and the most used models. This needs the Account Analytics: Read permission on the token; without it, the summary says so.

Costs

Workers AI bills by usage, in neurons. Cloudflare’s free plan includes 10,000 neurons a day, and larger models use far more per reply than small ones. Check your usage in the summary in Inference engines, or in the Cloudflare dashboard.

DigitalOcean Serverless Inference

Uses the models DigitalOcean hosts, through its OpenAI-compatible API. There is nothing to deploy, and one key is all it needs.

Field	What to enter
Model access key	Create one in the DigitalOcean control panel under Inference Engine → Serverless Inference → Model Access Keys

What works

• Many models. Open models such as Llama, DeepSeek, Qwen, Mistral, Gemma, GLM, Kimi and Nemotron, and commercial ones such as OpenAI’s GPT and Anthropic’s Claude. Each chat can use a different one.
• Streaming chat, with the generation settings DigitalOcean supports: temperature, Top-P, reply length and seed. Top-K and repeat penalty are not sent.
• Knowledge base. DigitalOcean’s embedding models, such as bge-m3 (many languages), gte-large-en-v1.5 and qwen3-embedding-0.6b, work for document search. Pick one as the Default embedding model under Admin settings → System → Default models. The same notes on switching embedding models apply.
• Chat titles and other short internal requests.

Not every listed model works on your plan

DigitalOcean lists all of its models, including ones your account or plan cannot use yet, such as some commercial models. Choosing one of those gives an error when the chat replies. Pick another model, or check which models your plan includes in the DigitalOcean control panel.

pAIring leaves out the models that are not for text chat, such as image, video and speech models, so they never appear in the model list.

Summary in Inference engines

When DigitalOcean is the active engine, Admin settings → Inference engines shows its status, whether the key is active, and how many chat and embedding models it lists. DigitalOcean has no usage API for this service, so check usage and costs in its control panel under Billing or Inference Engine → Serverless Inference.

OpenRouter

One key for hundreds of models from many providers, through OpenRouter’s OpenAI-compatible API. By default pAIring lists only the free models, so you can try it at no cost.

Field	What to enter
API key	Create one on openrouter.ai under Settings → Keys
Models to list	free (the default, if left empty) lists only the free models. all lists every model, and paid ones use your credits

What works

• Many models. Open and commercial models from many providers. Each chat can use a different one. Free models end in “(free)” on OpenRouter.
• Streaming chat, with your generation settings. Temperature, Top-P, Top-K, repeat penalty, reply length and seed are passed on; a model that lacks one ignores it.
• Knowledge base. OpenRouter’s embedding models work for document search. Few or none of them are free, so set Models to list to all to see them. The same notes on switching embedding models apply.
• Chat titles and other short internal requests.

pAIring leaves out the models that also make audio or images, such as music models, so only text chat models appear.

Privacy of free models

The providers of some free models may log your prompts or use them for training. Avoid sending private data to free models, or check each model’s data policy on OpenRouter.

Summary in Inference engines

When OpenRouter is the active engine, Admin settings → Inference engines shows whether the key is active, which models are listed, free-model requests used today against your daily limit, credits used today and this month, and whether the account has bought credits. Free use has a daily request limit, which is higher once you have bought some credits.

Google Gemini

Uses Google’s Gemini and Gemma models through the Gemini API, with the free tier of Google AI Studio. No card is needed to start.

Field	What to enter
API key	Create one in Google AI Studio (aistudio.google.com) under Get API key

What works

• Gemini and Gemma models, such as Gemini Flash, Flash-Lite and Pro. Each chat can use a different one.
• Streaming chat, with temperature and Top-P. Top-K, repeat penalty, seed and reply length are not sent: Google rejects a seed, and Gemini models think before they answer, so a reply-length limit could leave the reply empty.
• Knowledge base. gemini-embedding-001 works for document search, in many languages including Hebrew. The same notes on switching embedding models apply.
• Chat titles and other short internal requests.

pAIring leaves out the models that are not for text chat, such as image, speech, video, music and long-running research models.

Not every listed model works

Google lists some models that fail for some accounts: an older model may answer that it is no longer available to new users, and some Gemma models return an internal error from Google’s side. When busy, Google may also ask you to try again later. If a model fails, pick another one, such as a Gemini Flash model.

Privacy of the free tier

On the free tier, Google may use your prompts and replies to improve its products, and people may review them. In the EEA, the UK and Switzerland, Google allows only the paid tier. Use a paid (billing-enabled) key for private data.

Summary in Inference engines

When Gemini is the active engine, Admin settings → Inference engines shows its status, whether the key works, and how many chat and embedding models it lists. Google has no usage API for this, so check your rate limits and usage in Google AI Studio under Usage.

Groq

Very fast replies from hosted open models, through Groq’s OpenAI-compatible API. The free tier needs no card and has rate limits per model.

Field	What to enter
API key	Create one at console.groq.com under API Keys

What works

• Open models, such as Llama, Qwen, GPT-OSS and Kimi. Each chat can use a different one.
• Streaming chat, with temperature, Top-P and seed. Top-K, repeat penalty and reply length are not sent; some Groq models think before they answer, so a reply-length limit could leave the reply empty.
• Chat titles and other short internal requests.

pAIring leaves out the speech-to-text, text-to-speech and safety-classifier models, so only text chat models appear.

Chat only

Groq has no embedding models, and the knowledge base always uses the active engine’s embeddings. So while Groq is the active engine, the knowledge base is unavailable; pick another engine to use it.

Summary in Inference engines

When Groq is the active engine, Admin settings → Inference engines shows its status, whether the key works, and how many chat models it lists. Groq has no usage API, so check your limits and usage at console.groq.com under Settings → Limits and Usage.

xAI Grok

Uses xAI’s Grok models through xAI’s OpenAI-compatible API. Usage is billed per use to your xAI account. Not to be confused with Groq, a different company that hosts open models.

Field	What to enter
API key	Create one in the xAI console (console.x.ai) under API keys. It starts with xai-

What works

• Grok chat models, listed live from xAI. Image-generation, video and voice models are left out.
• Streaming chat with temperature, Top-P and seed.
• Chat titles and other short internal requests.
• Not yet: the knowledge base (xAI has no embedding models), images and tools.

xAI recommends its newer Responses API for new work, and keeps the Chat Completions API that pAIring uses documented and supported. It covers plain text chat, which is what pAIring’s connectors use for now.

Summary in Inference engines

When xAI is the active engine, Admin settings → Inference engines shows its status, whether the key works, and how many chat models it lists. Check your usage and credits in the xAI console.

NVIDIA NIM

Uses the models in NVIDIA’s API catalog on build.nvidia.com, through its OpenAI-compatible API. A free developer account comes with credits and a rate limit of about 40 requests a minute.

Field	What to enter
API key	On build.nvidia.com, open any model and press Get API Key. It starts with nvapi-

What works

• 100+ models, such as Llama, Nemotron, Mistral, DeepSeek and Qwen. Each chat can use a different one.
• Streaming chat, with temperature, Top-P and seed. A reply length is always sent: some NIM models stop very early without one, so pAIring asks for up to 4,096 tokens, and at least 1,024 for short requests such as chat titles.
• Knowledge base. NVIDIA’s embedding models, such as nvidia/llama-3.2-nv-embedqa-1b-v2, work for document search. Pick one as the Default embedding model. The same notes on switching embedding models apply.
• Chat titles and other short internal requests.

pAIring lists only text chat and embedding models. Safety, reranking, speech, vision, image, video and science models in the catalog are left out.

Credits

The free credits are meant for trying models, not for production use. When they or the rate limit run out, chats show an error that says so. Check your credits on build.nvidia.com.

Summary in Inference engines

When NVIDIA NIM is the active engine, Admin settings → Inference engines shows its status, whether the key works, and how many chat and embedding models it lists. NVIDIA has no usage API for this, so check your remaining credits on build.nvidia.com.

Cohere

Uses Cohere’s Command chat models and Embed models through Cohere’s OpenAI-compatible API. A free trial key is enough to start.

Field	What to enter
API key	Create one in the Cohere dashboard (dashboard.cohere.com) under API keys. A trial key is free

What works

• Command models for chat, strong at answering from documents, in many languages.
• Streaming chat, with temperature, Top-P and seed.
• Knowledge base. Cohere’s Embed models work for document search. Pick one as the Default embedding model. The same notes on switching embedding models apply.
• Chat titles and other short internal requests.

pAIring reads the model list from Cohere itself, which says what each model is for. Deprecated models and models only for reranking are left out.

Trial key limits

A trial key allows about 1,000 calls a month and a few a minute, and is meant for trying, not production use. When the limit is reached, chats show an error that says so. A production key removes the trial limits.

Summary in Inference engines

When Cohere is the active engine, Admin settings → Inference engines shows its status, whether the key works, and how many chat and embedding models it lists. Check your calls and limits in the Cohere dashboard.

Azure OpenAI

Uses the models you deployed on your own Azure OpenAI or Microsoft Foundry resource, through Azure’s v1 API. That covers OpenAI models such as GPT-4.1 and GPT-5, and Foundry models such as DeepSeek, Grok, Llama and Mistral. Usage is billed to your Azure subscription, and your prompts stay in your resource and its region.

Field	What to enter
Endpoint	Your resource’s endpoint from Keys and Endpoint in the Azure portal, such as https://my-resource.openai.azure.com or https://my-resource.services.ai.azure.com. Pasting it with /openai/v1 at the end also works
API key	KEY 1 or KEY 2 from the same page
Chat deployments	Your deployment names, comma separated, such as gpt-4.1-mini, DeepSeek-V3. These are the names you gave when deploying, not the model names
Embedding deployments	Optional: embedding deployments for the knowledge base, such as text-embedding-3-small

Deploy at least one model in the Azure portal or Microsoft Foundry first. Azure’s API does not list your deployments, so pAIring offers exactly the deployments you enter. Test connection checks the endpoint and key.

What works

• Streaming chat with temperature, Top-P and seed.
• Reasoning models (the o-series and GPT-5) reject temperature, Top-P and seed. pAIring notices this on the first reply, retries without them, and remembers it for that deployment.
• Knowledge base. Pick an embedding deployment as the Default embedding model. The same notes on switching embedding models apply.
• Chat titles and other short internal requests.

When a reply fails, the error says why: a wrong key, a deployment name that doesn’t exist, or a rate or quota limit reached. To raise a deployment’s limit, change its tokens-per-minute setting in the Azure portal.

Summary in Inference engines

When Azure OpenAI is the active engine, Admin settings → Inference engines shows its status, whether the key works, and how many chat and embedding deployments are set. Usage and cost are in the Azure portal, under your resource’s Metrics and Cost Management.

Google Vertex AI

Uses Gemini and Model Garden partner models (such as Llama, DeepSeek and Qwen) on your own Google Cloud project, through Vertex AI’s OpenAI-compatible endpoint. Usage is billed to the project, and its region and data controls apply. This is separate from the Google Gemini connector, which uses a Google AI Studio API key instead.

Vertex AI signs in with a service-account key, not an API key. pAIring exchanges it for a short-lived access token and renews it by itself before it runs out.

Field	What to enter
Service-account key (JSON)	In Google Cloud, IAM → Service accounts: create one with the Vertex AI User role, then Keys → Add key → JSON. Paste the whole file. It is stored encrypted
Project ID	Optional. Leave blank to use the project the key belongs to
Location	Optional. A region such as us-central1 or europe-west4, or global. Blank means us-central1
Models	Comma separated, as Vertex names them, such as google/gemini-2.5-flash, google/gemini-2.5-pro

The Vertex AI API must be enabled in the project, and partner models enabled in Model Garden. Test connection signs in and asks your first model for a one-token reply, so it checks the key, project, location and model together.

What works

• Streaming chat with temperature, Top-P and seed.
• Chat titles and other short internal requests.
• Not yet: the knowledge base (embeddings), images and tools.

When a reply fails, the error says why: the API not enabled or a missing role, a model name that doesn’t exist in that location, or a quota reached.

Summary in Inference engines

When Vertex AI is the active engine, Admin settings → Inference engines shows its status, whether the key still signs in, and how many models are set. Usage and cost are in the Google Cloud console, under the Vertex AI dashboard and Billing.

Amazon Bedrock

Uses models on your own AWS account through Bedrock’s OpenAI-compatible Chat Completions API, on its recommended bedrock-runtime endpoint. That includes OpenAI’s GPT-OSS, Anthropic Claude, Meta Llama, Mistral, DeepSeek, Qwen and more. Usage is billed to your AWS account.

You can sign in two ways:

• Bedrock API key. The quickest way to try it. Create a long-term key in the Bedrock console under API keys, in the Region you will use. AWS recommends long-term keys for exploration only.
• IAM access key. Recommended for long-term use. An IAM user’s access key ID and secret access key, with bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream. pAIring signs every request with it (AWS Signature Version 4), so nothing long-lived is sent.

Field	What to enter
Region	The AWS Region, such as us-east-1 or eu-central-1
Bedrock API key	Either a Bedrock API key…
Access key ID and Secret access key	…or an IAM access key. Fill in one way, not both
Models	Model IDs or inference profile IDs, comma separated, such as openai.gpt-oss-120b-1:0, us.anthropic.claude-sonnet-4-6

Use models whose model card in the Bedrock docs lists Chat Completions on bedrock-runtime. Bedrock’s API does not list your models, so pAIring offers exactly the ones you enter. Test connection asks your first model for a one-token reply, so it checks the Region, credentials, permissions and model together.

What works

• Streaming chat with temperature, Top-P and seed. Some models refuse some of these, such as newer Claude models, which take temperature or Top-P but not both. pAIring retries without the refused ones and remembers that for the model.
• Chat titles and other short internal requests.
• Not yet: the knowledge base (embeddings), images and tools.

When a reply fails, the error says why: a key made in another Region, missing permissions, a model not available here, or a quota reached.

Summary in Inference engines

When Bedrock is the active engine, Admin settings → Inference engines shows its status and how many models are set. Bedrock has no free call to check the connection with, so the status shows whether the connector is set up and tested. Usage and cost are in the AWS console, under Bedrock’s metrics and Billing.

RunPod Serverless

Talks to your own RunPod endpoint that runs the vLLM worker, which offers an OpenAI-compatible API. pAIring calls your endpoint directly. Nothing extra runs locally. RunPod bills by the time your endpoint runs; there is no free tier.

Field	What to enter
Endpoint ID	From your endpoint’s URL in the RunPod console
API key	A RunPod API key with access to that endpoint
Model name	The model your endpoint serves. Used when the model list cannot be read live

Before you start, deploy the endpoint in RunPod with the vLLM worker template and the model you want.

What works

• Streaming chat, with your generation settings. Temperature, Top-P, Top-K, repeat penalty and reply length are passed to vLLM.
• Chat titles and other short internal requests.

Chat only

RunPod Serverless does not do embeddings here, and the knowledge base always uses the active engine’s embeddings. So while RunPod is the active engine, the knowledge base is unavailable.

What does not work yet, on every connector

• Images and tools. Messages are sent as text only, so the chat’s Tools button is hidden while a connector is the active engine.
• Model management. Models are managed by the provider, so the Hugging Face browser is hidden while a connector is active.