Qwen3.8 27B

“So far with 3.8 i just need to check the end result...”

“q3 xl unsloth. 150k context. hatch 512 and kv q4. Its solid. review luna and sol plans and catch gaps confirmed later by them”

“it can follow my coding conventions and even document the architecture the right way without much hand holding. No other model that fit on my 16GB VRAM does as well.”

“Qwen3.8 27b is a very capable model, and there isn't much to gain for pursuing larger models despite the significant investment required.”

“4090 256k speed is 45tok/s, dropping to 128k would probably get around 55”

“Qwen3.8 27B have very similar behavior to Minimax M3 (thinks a lot, check everything, and like to continue working on its own for long session until it's done). It's like a slightly worse M3.”

“Qwen 3.8 27B is the real deal BUT remember to use froggeric template and/or medium reasoning”

“Qwen 3.8 27B Is extremely efficient when It comes to KV cache”

“they work very well, like very very well in a 5060 with 16GB”

“works very well with Pi out-of-the-box”

“it thinks forever and sometimes in a loop”

“It can hallucinate with historical facts so dojlnt trust too much”

“qwen3.8 27b is the same workflow but sloooooweeeeeerrrrrrr”

“it had hallucinated a key detail”

“It feels like it can never figure out how to update memory”

“Completely unusable last time I tried it. You got hit with rate limits after the first few minutes of using it. It's fast but cannot sustain its claimed speed before it immediately hits its rate limit. What good is it if you can't finish a”

“il a réfléchi 15 min et codé 2 min”

“in qwen website nothing models 3.8-27B”

“It's not perfect, not going to lie, however it's usable if you can give direct instructions”

“i see the memory use growing during tool use, with qwen3.8 27B it certainly plays a role”

“falls apart on multi-step agentic tasks. Tried it on a refactor that needed search plus edit across files and by the third tool call it was citing functions it had already overwritten.”

“temp 1 will suck no matter the quant, 0.7 is the temp to use, idk why they reccomended temp 1”

“Cost-performance is far worse than API”

“For actual work, if it's even slightly complex it'll all fall apart. My local setup can only be used for pure entertainment - if I want to get work done I still have to pull cloud models”