The authors describe Qwen3.8-Flash-Next, a sparse mixture-of-experts model containing 125 billion total parameters with 6 billion activated per token. ^[1] The model leads the 397-billion parameter A17B predecessor on eight of fourteen pre-training benchmarks while trailing on the remaining six by at most 2.6 points. ^[2] Qwen3.8-Next achieves this performance at one-third the activated parameters, one-third the training tokens, and roughly one-ninth the training FLOPs compared to the predecessor. ^[3] Token mixing employs a layer-wise hybrid of Gated DeltaNet and global attention, utilizing one full-attention layer in every four layers. ^[4] A Lemmy post compiled a list of 29 research and engineering directions for large language models, each with a brief description of its purpose and potential benefits. ^[5] Latent reasoning performs reasoning in continuous vector representations instead of discrete token sequences, allowing more information to pass between reasoning steps. ^[6] Linear-attention and hybrid architectures replace or combine standard attention with fixed-state mechanisms to reduce inference cost and KV-cache memory. ^[7] The post notes that linear attention is most advantageous at very long contexts, while retaining some full-attention layers can preserve exact recall, making the optimal layer mix a central design decision. ^[8]