Qwen3.8 Flash

“Damn son”

“i did notice it over engineers as well...”

“the final result was perfect: no bugs, and everything magically implemented and working. The time investment is worth it with this model”

“hallucinates in the middle”

“It's fine for pure execution, but lacking in planning. The intuition for handling business problems isn't great, not smart enough.”

“enabling cache makes it get about 20 tokens/s with a r9700 / rx9070 with 256k context / kv on f16 & 128gb ddr4”

“It is a bit smaller while losing hardly any precision”

“I really love this model so far”

“The slowness is unbearable, I won't use it anymore.”

“Qwen3.8 flash is like an idiot, either stops working halfway or has various hallucinations and tool call failures.”

“Running this model seems to require 128g memory plus 16g vram. Now 128g memory also costs over 10,000 yuan.”

“thinking default is x-high super annoying and was even hallucinating/looping quite often”

“qwen3.8-flash >＝ dsv4f > glm-5.3-flash, but Qwen really does go into full-on 'exam-taking' mode”