So ur here to learn about Qwen, so break it short, the Qwen family Iz the most recommended family for low ram users due to its brilliant architecture that keeps u up all night
- It's honestly a pretty good starting point for a lot of users simply because of how broad a net it casts i recommend for anyone who wants to skip research, all rounder pretty good
- The Qwen and her royal incestuous family, we have Qwen2.5 being the main outreach model, but it's garbage it came out it 2025 jan, that's prehistoric
- The latest model like Qwen3.5-3.7 are kinda good but they lost the plot of serving low end RAM users they use fancy terms like gated attention which just means how much they can forget to let new info get saved about yours truly, deltanet is the part that handles new info more directly, and lastly the enemy the affairs of the Queen the MoE
- The scam, MoE sounds good, sparse attention Yaaay, or is it, sparse attention just means they fire parts of the model, like how u don't use ur stomach to walk,the scam is just because it's firing some "experts" doesn't mean the non used ones are not cold started, they will use some background ram, also most experts would never be used, so u can just get grafted variations,
- Always define what u want from Qwen not the other way around, she has a hot and stubborn temper and body 👅, so if u catch her cheating she won't admit it normally unless u get a variation that's a bit more sub
- My advice to a beginner with a potato try qwen3.5 hyper quantised, or qwen2.5 to get used of the queen's tantrums, then u can try going crazy, also start at 8k, and get ready to be gaslit more than ur dating life For masos: umm yeah graft it or lora train to reduce hallucinating For deep throats: try using Deepseek R1 distill with qwen instruct and Q4-5 quantisation or try the new IQ quantisation, or even splice it if u feel like Frankenstein
- Ask me questions or the French Qwen will haunt you
Top comments (0)