Qwen · model № 002
Qwen3.8-Flash-Next
A 125B model that moves 51B n-gram parameters into host memory and targets a ninth of the flagship training compute
total params
125B
+51B n-gram tables in RAM
active / token
6B
≈1/3 of its predecessor
training FLOPs
1/9×
vs the 397B flagship
design method
3-axis
quality · cost · stability
Chapters
- 1The 1/9 RecipeHow Qwen targets quality, training cost, and stability together, rather than optimizing one after the fact.
- 251B Params in Your RAMWhy 51B n-gram parameters live in host memory, how prefetching works, and where the tradeoff lands.
- 3The Micro-Block SpotlightHow a micro-block index narrows a large context before the expensive attention calculation begins.
- 4Four Lanes and a Toll BoothHow a four-lane residual path carries more memory through depth, and how each lane is gated.
- 5The Stability DividendWhy the optimizer and residual design make large training runs more stable, and what that stability saves.
- 6The Flash-Tier WarThree flash models make different tradeoffs between active parameters, memory traffic, and price.