Best Ollama Local Model To USE #362
Description
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.
Metadata
Metadata
Assignees
Labels
No labels
Activity
Chenglong-MS commentedon Jun 16, 2026
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 commentedon Jun 16, 2026
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 commentedon Jun 17, 2026
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 commentedon Jun 17, 2026
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:
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.