AI Development general thread.

I'm going to be spending a little time this weekend with codex using local LLMs with Ollama - anyone have any comment/strong views on the following models?

devstral-small-2:24b
qwen3-coder:30b
qwen3.6:27b
gemma4:31b
qwen3-vl:30b
 
quovadis said:
I'm going to be spending a little time this weekend with codex using local LLMs with Ollama - anyone have any comment/strong views on the following models?

devstral-small-2:24b
qwen3-coder:30b
qwen3.6:27b
gemma4:31b
qwen3-vl:30b
Click to expand...
I would strongly suggest startng with something non critical or a copy of, especially if you are switching model families. I had some strange behavior a while back when I played with it. Also some wierd things between codex and ollama when talking to non openai models ( it used some form of bridge ). Linux based system though which seems to be tail end of openai dev tool support.
 
Kosmik said:
I would strongly suggest startng with something non critical or a copy of, especially if you are switching model families. I had some strange behavior a while back when I played with it. Also some wierd things between codex and ollama when talking to non openai models ( it used some form of bridge ). Linux based system though which seems to be tail end of openai dev tool support.
Click to expand...
Sure, I wasn't planning to use it for any actual work. More evaluation of new from scratch, perhaps some tests modifying a static code base I already used when evaluating vanilla codex. This will be on a Mac Studio.
 
quovadis said:
Sure, I wasn't planning to use it for any actual work. More evaluation of new from scratch, perhaps some tests modifying a static code base I already used when evaluating vanilla codex. This will be on a Mac Studio.
Click to expand...
You can also test ollama cloud llms. One of my experiments is a little agent bot that I have written in Python using tiny llm and discord as a comm medium but it switches between a ollama cloud free high model to local depending on need as I'm running it on a respective potato.
 
quovadis said:
I'm going to be spending a little time this weekend with codex using local LLMs with Ollama - anyone have any comment/strong views on the following models?

devstral-small-2:24b
qwen3-coder:30b
qwen3.6:27b
gemma4:31b
qwen3-vl:30b
Click to expand...
They all suck.
 
Codex bugs apparently fixed last night. Also reminded people that the Code Review side also eats up your codex usage even though that wasn't clear initially.
 
quovadis said:
I'm going to be spending a little time this weekend with codex using local LLMs with Ollama - anyone have any comment/strong views on the following models?

devstral-small-2:24b
qwen3-coder:30b
qwen3.6:27b
gemma4:31b
qwen3-vl:30b
Click to expand...

Really depends on your setup and what specs you are running been impressed with Qwen adn Gemma but the smaller models, would recommend you try qwen3.6:27 and gemma4:31b.

The never models should provide big improvements over the older models.
 
L-Dog said:
Really depends on your setup and what specs you are running been impressed with Qwen adn Gemma but the smaller models, would recommend you try qwen3.6:27 and gemma4:31b.

The never models should provide big improvements over the older models.
Click to expand...
gemma4:26b actually performed decently for the limited testing I did.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X