FluentDBFluentDB
Download
AI models

Run FluentDB's AI locally with Ollama

Point FluentDB at your own Ollama server and the model runs on your Mac. No API key, no token bill, nothing leaves the machine.

Ollama runs open-weight models on your own hardware. FluentDB talks to it over the local HTTP API, so your schema, your prompts and your queries never go anywhere. No key to rotate, nothing metered, no account.

The catch is that local models are all over the place. One will write you a perfectly nice paragraph and then hand back SQL that doesn't parse, or call a tool with arguments it made up on the spot. So the list below is short, deliberately. A model gets on it after we've pointed it at real schemas and watched what it does.

Supported today

One model, tested properly

qwen3-coder

Supported

A 30B mixture-of-experts coder that only fires 3.3B parameters per token. Native tool calling, 256K of context, Apache 2.0. It's the best local model we've found for writing SQL against a real schema.

ollama pull qwen3-coder
Full name
Qwen3-Coder-30B-A3B-Instruct
Ollama tag
qwen3-coder (same as :30b)
Architecture
Mixture of Experts
Parameters
30.5B total, 3.3B active per token
Experts
128, with 8 active per token
Layers
48
Context window
262,144 tokens
Quantization
Q4_K_M
Download size
~19 GB (17.3 GiB)
Tool calling
Native, parsed by Ollama
Thinking mode
None. Instruct only
License
Apache 2.0
Released
July 2025

Where it falls short

In our testing the model doesn't cope well with a dead end. When it can't find the right answer it keeps going, and starts making things up rather than stopping. You'll see it most with Auto Run on and AI Read Access off. The query runs, but the model never gets the results back, so it has no way of knowing whether anything worked. Instead of stopping to ask you, it writes another query. Then another. We're working on it.

Set it up in five steps

Ollama does the hosting. FluentDB just talks to it on localhost.

  1. 1Install Ollama from ollama.com, or with brew install ollama.
  2. 2Pull the model with ollama pull qwen3-coder. It is about 19 GB, so give it a minute.
  3. 3Make sure Ollama is running. It serves on http://localhost:11434 by default.
  4. 4In FluentDB, open Set up AI, choose Ollama (local), and check the server URL matches yours.
  5. 5Pick qwen3-coder from the supported models list, then save.

Why the list is short

Every model on this page has been pointed at real schemas, making real tool calls, on a real Mac. If one you like is missing, it's usually for one of two reasons.

It can't do something we need

The co-pilot works in a loop: read the schema, call a tool, run something, read the result, decide what's next. Plenty of local models that write beautifully can't call a tool reliably, and a few can't call one at all. That's a hard stop, and no amount of good prose gets around it.

It gets the SQL wrong

Some models call every tool perfectly and still write SQL that quietly returns the wrong number. A query that errors is annoying for about five seconds. A query that looks right and is wrong can sit in a report for a month before anyone notices. Those ones don't ship.

We'd rather ship three models that work than thirty that sometimes do.

Ollama FAQ

Which Ollama models does FluentDB support?

Right now, qwen3-coder. Ollama will run hundreds of models, but most of them can't handle schema-aware SQL and reliable tool calling well enough for us to put our name on them. We add models as they pass testing.

Does my data leave my Mac when I use Ollama?

No. Ollama runs the model on your own hardware and FluentDB talks to it locally. No API key, and no request to us or to any model vendor.

Do I need an API key to use Ollama?

No, and that's most of the appeal. You need the Ollama app running and the model pulled. Nothing is metered and there's no account.

What hardware do I need for qwen3-coder?

The download is about 19 GB, and those weights stay in memory the whole time the model is loaded, before you add any context. A 32GB Mac handles it comfortably. 24GB works, but you'll feel the squeeze once the context grows.

Is qwen3-coder as good as Claude or GPT?

For everyday queries, it's good. For the hard ones it's behind, and we'd rather just say so. On Spider2, a demanding text-to-SQL benchmark, the 30B scores 21.4 where frontier models get to about 31.1. Use it for daily work, switch to an API model when the question is hard.

Can I suggest a model?

Yes. Send it through the support page and we will look at it. We test everything we ship, so a suggestion puts a model in the queue rather than guaranteeing it lands.

Stop wrestling
with your database.

Download FluentDB and run your first AI-assisted query in 60 seconds.

Download for Mac