Gpt 4o-mini vs models

Posted by Osama_Saba@reddit | LocalLLaMA | View on Reddit | 8 comments

What size of the Qwen-3 model is like the gpt-4o mini? In terms of not being stupid

Reply to Post

8 Comments

[-]

MKU64@reddit

Pretty sure none is like GPT 4o-mini tbh. At least in my use case which is UI Prototyping. I consider GPT 4o-Mini incredibly underrated and it’s mostly because it hasn’t been updated in a long time. Give it knowledge of instruction following protocols of the present and current date knowledge and you have a competing model to a lot of others out there.

[-]

netixc1@reddit

This model is cheap on openrouter and is good with tools, i tend to use it more then local llm's cuz they always F things upp

[-]

MKU64@reddit

Exactly and it totally does the job way better at coding than 4.1 Nano which they wanted to use it to compete? Not even close

[-]

compiler-fucker69@reddit

[https://dubesor.de/benchtable](https://dubesor.de/benchtable) use this site much closer for my usecase ngl other than that for hallucination the results are grounded in reality and yeah private benchmark no contamination , do not use vectera ones for hallucination most people say the benchmark is less than 1k tokens to test hallucination and forgetfulness for my usecase i have not found a model yet will update once i am done making my own benchmark let's hope it gets done

[-]

lly0571@reddit

Maybe Qwen3-14B or 30B-A3B. I consider 4o-mini as something close to Qwen2.5-32B or Gemma3-27B personally.

[-]

Osama_Saba@reddit (OP)

But qwen 14b is smarter than 2.5B like my dog is smarter than a fruit fly

[-]

No_Expert1801@reddit

Qwen 3 14b maybe?

[-]

sammoga123@reddit

I think a Microsoft study came out a while ago that put parameters on private models, so, in theory, GPT-4o mini is 8b