Best Free AI Models for n8n (2026): Power AI Agents Without Paying for Tokens

Updated July 2026

Best Free AI Models for n8n (2026): Power AI Agents Without Paying for Tokens

The best free AI models for n8n in 2026 are Llama 3.3 70B on Groq (1,000 requests per day, no charge), Gemini 2.5 Flash on Google AI Studio (generous free tier, no credit card), and the ":free" models on OpenRouter (one key, dozens of models). All three expose OpenAI-compatible endpoints, which means n8n's AI Agent and LLM nodes can use them today with a standard OpenAI credential and a custom base URL.

That is the short answer. The longer answer involves rate limits, retry logic, and the point where a free tier quietly becomes the most expensive part of your stack. If you are building n8n workflows that summarize emails, qualify leads, or run AI agents, this guide shows you exactly which free model to pick, how to wire it in, and when free stops being the right call.

We maintain a live directory of 297 free AI models, community-ranked with rate limits and the catch behind each one, so the specifics below stay current even as providers change their terms.

What are the best free AI models for n8n?

Five providers stand out for n8n builders in 2026. Each one gives you a real model, a real API key, and zero token cost, with different trade-offs on speed, limits, and reliability.

For a broader directory that covers paid models alongside the free tiers, see our model directory.

Groq is the default recommendation. Llama 3.3 70B is smart enough to drive an n8n AI Agent node with tool calling, and 1,000 requests per day covers most internal automations. The endpoint at https://api.groq.com/openai/v1 is fully OpenAI-compatible, so setup takes two minutes.

Google AI Studio is the pick when your workflow needs long documents or images. Gemini 2.5 Flash has one of the most generous free tiers on the market and does not ask for a credit card, which makes it the lowest-friction starting point of the five.

OpenRouter is the pick when you want options. One API key at https://openrouter.ai/api/v1 reaches every model tagged ":free", roughly 20 requests per minute and between 50 and 1,000 requests per day depending on the model. It is the easiest way to A/B test models inside the same n8n workflow without juggling credentials.

Cerebras and Mistral Codestral are specialists. Cerebras serves Llama 3.3 70B at the fastest inference speeds available on a free tier, which matters when a human is waiting on the workflow output. Codestral is free with a Mistral key at 30 requests per minute and outperforms general models on code tasks.

Key takeaway: start with Llama 3.3 70B on Groq for agents, Gemini 2.5 Flash for long context, and OpenRouter when you want to test several models with one key. All three cost nothing and plug into n8n through the standard OpenAI credential.

How do you connect a free AI model to n8n?

Every provider above speaks the OpenAI API format. n8n's AI Agent node and LLM Chain nodes accept any OpenAI-compatible endpoint, so you never need a community node or a custom HTTP request to use them.

Here is the exact setup, using Groq as the example:

1. Create a free account at the provider's console and generate an API key. For Groq that is console.groq.com; no card required.
2. In n8n, open Credentials, create a new OpenAI credential, and paste the key into the API Key field.
3. In the same credential, set the Base URL to the provider's endpoint: https://api.groq.com/openai/v1 for Groq, https://openrouter.ai/api/v1 for OpenRouter, and the equivalent from each provider's docs for the rest.
4. Add an AI Agent or OpenAI Chat Model node to your workflow and select the new credential.
5. Type the model name exactly as the provider lists it, for example llama-3.3-70b-versatile on Groq. Free endpoints often will not appear in n8n's model dropdown, so manual entry is normal.
6. Run a test execution and confirm you get a completion back before wiring it into anything downstream.

The same recipe works for Google AI Studio through its OpenAI-compatible endpoint, for Cerebras, and for Mistral. Swap the base URL, the key, and the model name; nothing else in your workflow changes.

Two practical tips from production builds. First, store one credential per provider and name it clearly (for example "Groq free tier"), because a month from now you will not remember which key hits which limit. Second, set the AI Agent node's max iterations conservatively; a looping agent can burn a 1,000-request daily quota in a single bad execution.

Which free model should you pick for each workflow type?

Match the model to the job instead of picking one and forcing it everywhere.

Lead qualification and email triage: Groq's Llama 3.3 70B. Classification and extraction are comfortably inside its capability, and 30 RPM absorbs bursty inbox traffic.

Document summarization and RAG: Gemini 2.5 Flash. Long context handles full PDFs and transcripts without the chunking gymnastics smaller context windows force on you.

Chat-facing agents where speed is felt: Cerebras. When a customer or teammate is watching the response render, its inference speed is the difference between snappy and abandoned.

Code generation inside workflows: Codestral. If your workflow writes SQL, transforms JSON in a Code node, or drafts scripts, a code-tuned model beats a generalist at the same price of zero.

Experimentation and fallbacks: OpenRouter. Because dozens of ":free" models sit behind one credential, you can chain a second model as a fallback branch when the first returns an error.

Our free models directory tracks all 297 of these with current limits, so check it before committing a workflow to any single provider.

When do free tiers break in production n8n workflows?

Free tiers are engineered for developers testing, not businesses operating. Three failure modes show up on almost every production n8n deployment that runs on free models.

Rate limits collide with real volume. Do the math before you ship. Groq's 1,000 requests per day sounds like a lot until you spread it across 24 hours: that is roughly one workflow run every 90 seconds, sustained. A webhook-triggered workflow that fires on every form submission, every inbound email, and every CRM update will hit the ceiling by lunch. OpenRouter's ":free" models can be as low as 50 requests per day. If a single workflow execution makes three LLM calls (agent reasoning, tool call, final answer), divide your daily budget by three again.

Reliability is not guaranteed. Free endpoints get deprioritized when providers are under load, and models get renamed, upgraded, or removed with little notice. Production workflows need retry with exponential backoff on 429 responses, an error branch that does something sensible when the model is down, and ideally a second provider wired in as a fallback. n8n gives you the error workflow and retry settings to do this, but someone has to build and test them.

Data terms differ on free tiers. Several providers apply different data usage and retention terms to free usage than to paid usage, and some reserve rights to review or use free-tier prompts to improve their services. Read the provider's data policy before routing customer emails, contracts, or anything regulated through a free endpoint. For internal, non-sensitive automation this is usually acceptable; for customer PII it often is not.

None of this means free models are a toy. It means they are a capacity tier, and you have to engineer around the edges of that tier the same way you would engineer around any other quota.

Key takeaway: free tiers fail on volume math, not model quality. Count LLM calls per execution, multiply by daily runs, and compare against the provider's cap before you ship. If the numbers are close, build retry, backoff, and a fallback provider from day one.

Can you self-host a free model for n8n instead?

Yes, and for some teams it is the better answer. Open-weight models like DeepSeek V3.2 and GLM 4.6 can run on your own hardware through Ollama (simplest) or vLLM (fastest for concurrent load). Both expose OpenAI-compatible endpoints, so the n8n setup is identical to the cloud recipe above: OpenAI credential, custom base URL pointing at your server, model name.

Self-hosting removes the three production problems in one move. There are no request-per-day caps, no third-party data exposure, and no surprise deprecations. Your prompts never leave your infrastructure, which is often the deciding factor for legal, healthcare, and finance workflows.

The trade is that "zero token cost" is not zero cost. Large open-weight models need serious GPU memory to run well, someone has to keep the inference server patched and monitored, and throughput planning becomes your problem instead of the provider's. A reasonable middle path many teams use: free cloud APIs for low-sensitivity workflows, a self-hosted model for anything touching customer data.

If you already self-host n8n, adding Ollama on the same network is a natural extension, and the two together give you a fully private automation stack.

When should you stop duct-taping free tiers and hire an expert?

There is a predictable arc here. The free model works in testing. The workflow ships. Volume grows, 429 errors appear, someone adds a Wait node, then a second API key, then a spreadsheet tracking which key is burned out. At that point the business is spending real engineering hours to avoid a token bill that would have been smaller.

You have outgrown the DIY stage when any of these are true:

• A workflow that touches revenue (lead routing, quoting, customer replies) depends on a free tier staying up.
• You are rotating API keys or providers manually to dodge rate limits.
• Nobody on the team knows what happens when the model returns an error at 2 a.m.
• Customer data is flowing through endpoints whose data terms nobody has read.

This is exactly the work we do. AY Automate designs and builds production n8n systems: multi-provider failover, retry and backoff done properly, self-hosted models where data demands it, and volume math done before launch instead of after the outage. Our custom workflow automation service covers the full build, from architecture through monitoring, so your automations behave like infrastructure instead of experiments.

If you are still comparing options, our guide to the best n8n agencies gives an honest look at the market, including where we fit and where another shop might suit you better.

Where AY Automate fits

We are not one of the tools on this list. We place AI engineers who use tools like these to generate revenue for your business, that is the offer. For n8n specifically, that means an engineer who wires free or paid models into your workflows and keeps them running. See AI engineer placement or explore our related service.

FAQ

Can I use n8n AI agents for free?

Yes. n8n's AI Agent node works with any OpenAI-compatible endpoint, and providers like Groq, Google AI Studio, and OpenRouter offer free tiers that require no payment. Create a free API key, add it as an OpenAI credential in n8n with the provider's base URL, and the agent runs at zero token cost within the provider's rate limits.

What is the best free LLM for n8n?

Llama 3.3 70B on Groq is the best free LLM for most n8n workflows, with 1,000 requests per day and 30 requests per minute at no charge. Gemini 2.5 Flash on Google AI Studio is the better pick for long documents and multimodal steps. Check our live free models directory for current limits across all 297 options.

Do free AI models work for production n8n workflows?

They work for low-volume production, but they need engineering around them. Free tiers cap requests per day, can be deprioritized under load, and may apply looser data terms, so production workflows need retry logic, error branches, and a fallback provider. Above roughly one execution per minute sustained, plan for paid tiers or self-hosting.

How do I set a custom base URL in n8n?

Open the OpenAI credential in n8n and replace the default base URL with the provider's endpoint, such as https://api.groq.com/openai/v1 or https://openrouter.ai/api/v1. Then enter the provider's model name manually in the AI node, since free models rarely appear in the dropdown.

Which free model is best for coding tasks in n8n?

Mistral's Codestral, which is free with a Mistral API key at 30 requests per minute. It is tuned for code generation, so it outperforms general free models when your workflow writes SQL, transforms data in Code nodes, or drafts scripts.

Is self-hosting cheaper than free API tiers?

Not at first: free API tiers cost nothing, while self-hosting DeepSeek V3.2 or GLM 4.6 via Ollama or vLLM requires GPU hardware and maintenance time. Self-hosting wins when data privacy rules out third-party endpoints or when your volume exceeds every free cap, because token cost stays at zero regardless of scale.

Who can build a production-grade n8n AI system for my business?

An automation agency with real n8n production experience, covering failover, rate limit handling, and data compliance rather than just workflow canvases. AY Automate builds these systems end to end through our custom workflow automation service, and you can book a consultation to scope your use case and get honest volume math before anything gets built.

For related guides in this cluster: free AI models for coding covers editor integrations (Cursor, VS Code, Claude Code), free AI models for commercial use covers which licenses let you ship without restrictions, and best local LLM tools 2026 covers self-hosted inference for air-gapped or high-volume setups.

---

For the cost structure of the frontier Claude model when your n8n workflows outgrow free-tier limits, see the Fable 5 pricing breakdown.

Sources: Groq documentation, Google Gemini API documentation, OpenRouter documentation, Cerebras inference documentation, Mistral documentation