Qwen 3.8 Flash Next uses N-gram table to save backbone weights

antirez · x · 2026-08-30

antirez explains the N-gram mechanism in Qwen 3.8 Flash Next: at Layer 1, it takes the last two and three tokens to retrieve 16 small vectors from a 51B database. These are composed into a larger vector and gated to enrich the current token at Layer 2. This approach saves significant weights on the model backbone by offloading fixed associations like "New York" to the table.

Related event: antirez: Qwen 3.8 Flash Next's 51B N-gram Table Could Be Best Local Inference Option on 64GB Macs(7 posts)→

Original post →

More from Models

Models channel →