Please, ChatGPT is hallucinating models, even with web-search on.
Posted by Ok-Type-7663@reddit | LocalLLaMA | View on Reddit | 7 comments
[32B-A3B does not exist]()
[sorry for the post i posted before, i was testing how people in localllama react to fake models, EVEN jacek2023 responded!]()
Yeah, chatgpt is inventing a lot of models even with web search
LocalLLaMA-ModTeam@reddit
Rule 2 - cloud service
Edenar@reddit
This looks like free tier chatGPT which isnt great, maybe ask real humans ? or Scroll this sub. You can also look what is popular on huggingface.
right now qwen 3.6 ( both 35b a3 and 27b) is quite popular. Qwen 3.5 122b is also an option with your hardware. Google gemma 4 models are also good for their size. A bit outdated but still good are gpt oss 120b and 20b. If you have enough ram/vram maybe quantized minimax m2.7 is also an option, ...
Good luck
Top-Rub-4670@reddit
OP is complaining about GPT using the wrong number of parameters in the model name.
People on this sub do the same all. the. time. I've stopped correcting them, frankly. You can usually guess which model they're talking about anyway...
jacek2023@reddit
who are all these users called "Ok-something"?
Ok-Type-7663@reddit (OP)
reddit default
Technical-Earth-3254@reddit
Isn't GPT providing source markers anymore behind sentences/Information? Just looking at the sheer amount of emojis, I would assume either the chat is super long and it starts to degrade (massively), web search isn't turned on or your custom instructs are telling it do to something weird..
Technical-Earth-3254@reddit
Gave it a shot. Yeah, it's a little off (Qwen 3.5 27B instead of 3.6, also recommending Mistral Small from last year in addition to Devstral 2 small and also gpt oss 20b and R1 Distill 32B) but it's usable (but not perfect).