Best Free LLMs for Workflow Automation: Mistral, LLaMA, Gemma

You don’t need an OpenAI account to build AI-powered workflows. The open-model ecosystem has matured to the point where free models handle the most common automation tasks (summarisation, classification, extraction, generation) with quality that’s more than enough for most uses.

This article compares the best free LLMs for workflow automation and explains how to pick the right one. All of them run locally with Ollama, and smaller versions of several run directly in your browser, with no API costs.

Why free models are now a serious option

A couple of years ago, the gap between open and commercial models was big enough to matter for production automation. It has narrowed a lot. Models like Mistral 7B, Llama 3 8B and Gemma 3 do well on the kind of short, focused tasks that workflow AI steps perform.

None of these needs a frontier-class model. And most of these models are shrunk (“quantised”) before you download them, which is why a 7B model fits in about 4 GB:

The top free LLMs for workflow automation

Sizes below are the default 4-bit downloads in Ollama; check the Ollama library for current tags.

Mistral 7B

Mistral 7B is still one of the best all-round models at its size. It’s fast, follows instructions well and is good with several languages. For most automation tasks, especially classification and summarisation, it’s a reliable first choice.

• Size: about 4.1 GB
• Best for: general-purpose automation, multilingual content
• Weakness: less strong on complex multi-step reasoning
• Run with: ollama pull mistral

Llama 3 8B (Meta)

Meta’s Llama 3 8B is one of the strongest models in its class. It follows instructions precisely, produces clean structured output and handles JSON formatting reliably, which matters when a node has to extract data and pass it on.

• Size: about 4.7 GB
• Best for: structured output, data extraction, instruction following
• Weakness: a little slower than Mistral on CPU
• Run with: ollama pull llama3 (or llama3.1 for the newer version)

Gemma 3 (Google)

Google’s Gemma 3 punches above its weight, with good reasoning for its size and coherent structured output. It comes in several sizes: the 4B model (about 3.3 GB) is light and quick, and the 12B model (about 8 GB) is noticeably better on harder extraction if your machine has the memory.

• Size: about 3.3 GB (4B) or 8 GB (12B)
• Best for: reasoning, document analysis, structured extraction
• Weakness: the 12B needs much more memory than the 7B models
• Run with: ollama pull gemma3 (4B) or ollama pull gemma3:12b

Phi-3 Mini (Microsoft)

Phi-3 Mini is the smallest serious model here. At about 2.2 GB it runs on modest machines and answers faster than the others. If your workflow makes many AI calls in a row (classifying 50 items in a loop, say), its speed is worth the quality trade-off.

• Size: about 2.2 GB
• Best for: fast classification, light generation, smaller machines
• Weakness: weaker on complex tasks
• Run with: ollama pull phi3

Qwen 2.5 7B (Alibaba)

Qwen 2.5 is a strong contender that’s often overlooked. At 7B it competes with Llama 3 on instruction following and is particularly good at code and data extraction.

• Size: about 4.7 GB
• Best for: code, technical extraction, structured output
• Weakness: less widely used in automation setups than Mistral or Llama
• Run with: ollama pull qwen2.5

Comparison table

The ratings are our general guide, not benchmark results. The best test is your own inputs: see do you need a big AI model? Test small ones on your own inputs.

How to choose the right model

Bigger is not automatically better for a workflow step. Size buys quality and costs memory and speed:

Then pick by task:

• General summarisation or classification: start with Mistral 7B.
• Clean JSON from a node: Llama 3 8B follows formatting instructions most consistently.
• Complex reasoning or multi-document analysis: Gemma 3 12B, if your machine can handle it.
• Many calls in a loop: Phi-3 Mini, for throughput.
• Code or technical content: Qwen 2.5.

Using free LLMs as nodes in Agentic Workflow

Agentic Workflow has an Ollama model node: start Ollama with the extension allowed (OLLAMA_ORIGINS, see the Ollama guide), pick one of your models in the node, and plug it into an AI step like Basic LLM Chain, Information Extractor or Text Classifier. Models running in Ollama are also listed automatically in the assistant and agent model pickers. You can also skip Ollama entirely: the Web LLM node runs small models (Llama, Phi, Qwen, Gemma variants) directly in your browser.

You build the same browser-native automations (reading a page, AI processing, showing the result) at zero API cost: no subscription, no per-token billing, no data leaving your machine. The canvas, the triggers and the nodes are identical to a workflow with a cloud model. The only thing that changes is where the AI runs.

More on this: how to use Ollama in a workflow automation tool, how to choose, install and compare a local AI model, and node-based workflow automation beyond Zapier and n8n.

Try it