Best Ollama Local Model To USE

[MrAnizHamada]

I’ve tried a lot of the local models from Ollama, and pretty much none of them have been able to run Data Formulator for me. I was expecting better results since my laptop is definitely not low-end:

• OMEN 16

• Intel® Core™ i9-14900HX

• RTX 5070 Laptop GPU (8 GB VRAM)

• 32 GB RAM

Even with this setup, I still haven’t found a local LLM that works well with Data Formulator. It feels like either the models I’m trying aren’t the right ones, or there’s something about Data Formulator’s requirements that’s not clear for local setups.

Could you help with:

• A list of Ollama models that are actually known to work with Data Formulator?
• The minimum specs you’d recommend (especially VRAM)?
• Any specific settings (quantization, context length, Ollama options) that should be used?
• If 8 GB VRAM is basically too small for this, what’s the smallest model that’s expected to work?

I’m hoping to get a working local setup and would really appreciate any guidance or a clear “these models work, these don’t” list.

[Chenglong-MS]

That's a good point, let me run a quick evaluation and get it back to you tmr. Recently development has all been on cloud models, so we haven't run much local model testing.

[MrAnizHamada]

I’ve already tested several local models with Ollama and Data Formulator, but all of them have problems. The models I tried are:

• deepseek-r1:7b

• deepseek-r1:8b

• deepseek-r1:14b

• qwen2.5-coder:7b

• qwen2.5-coder:14b

• gemma4:12b

• gemma3:12b

With all of these, the main issues are:

They take a very long time to “think” (generate responses).

At the end, I consistently get this error:

(data-agent) Unknown action: {{actionType}}

This makes it impossible to use Data Formulator reliably with local LLMs on my setup.

I really appreciate your prompt response to this issue. I’m waiting for your guidance and feedback tomorrow.

Thanks in advance.

[Chenglong-MS]

I see, let me do a testing sweep, I suspect 32b model is kinda required. i'll compare against gpt-5-mini, and hopefully from our collab, we can create a table to provide detailed guidance.

[Chenglong-MS]

Did some evaluation based on the https://github.com/rfordatascience/tidytuesday data and spider 2.0 datasets (45 data / question pairs), and use gpt-5.5 as the grader.

The current agent protocol is bad for anything outside of gpt-5.5, thus I built a mini-agent where the model can generate charts / responses from pure json results. This is the mini-agent eval results:

Tier	Model	Success (mini-agent)	Quality	Verdict
Reference	gpt-5.5 / gpt-5-mini	100%	89-91	ceiling
Best open	qwen3.6:27b	100%	87/85	matches refs, 1x80GB GPU
Strong open	gpt-oss:120b	96%	82-84	strong 2nd (MoE, fast)
Mixed	qwen3.6:35b	91%	82-84	inspection-dependent
Mid	qwen2.5:72b / coder:32b	80%	74-80	usable, mid
Single-table only	gpt-oss:20b	47%	82-89	great quality, low reliability
Weak	qwen2.5-coder:7b, llama3.1:8b	31-64%	67-75	unreliable on joins
Non-viable	llama3.2:1b	0%	-	protocol collapse

Seems qwen3.6 is overall the best one. I'll do a bit deepdive, and push the mini-agent to the dev branch later today.