I was trying Qwen 3.6-27b and Gemma4 in a siomple web chat. Asked them both a qn like 'recommend the best llm for a 5060ti' and was suprised when they both replied 'user is asking about a card that doesn't exist'.

I then saw their knowledge cutoff was early 2025, hence why. But tech advances so fast, in that 1yr+ there have been tons of changes in languages, frameworks, best practices and tech, not to mention AI.

of course the model could use MCP etc to do a web search, but its pretraining is still using 1yr+ older knowledge.

Found that suprisising and probably explains a few things, but its not something widely known I think