Qwen3.8 Flash Next

“Prompt prosessing changed a lot at lower pcie bandwidth while token generation was influenced a lot less”

“This is a frontier-class model”

“Qwen3.8 flash next you can probably do a Q4 or maybe even Q5, but you'll actually need 64 GB system RAM because of the 51B n-gram in that model.”

“It reasons well, it covers 90% of what frontier models used to do for me.”

“king of MOEs right now”

“Qwen3.8 flash next's implementation is very unstable in many inference engine. Give it more time.”

“Switched to it from DS4F for RAG summarizer, quality is definitely up, no garble. vLLM on sparks, hybrid NVFP4+FP8 checkpoint.”

“qwen3.8-flash-next's thinking is particularly long at xhigh level, way longer than glm-5.3-flash's max level”

“doing a really good job writing creative stories”

“Qwen 3.8 Flash Next is a big step in this direction”

“pretty impressed with what it can do with novel prompts”

“using nvfp4 radix quant the quality suffered compared to deepseek”

“I have been working with Qwen3.8 Flash Next, and while the level of speech is not amazing, I haven't seen anything like this in my custom harness.”

“fits on a 256GB system in Q8”