DeepSeek V3.2 Speciale slips on dense attention

This is less a model story than a serving-stack warning: DeepSeek’s sparse-attention design appears to matter materially once tasks get hard enough.

• –The benchmark suggests the dense fallback preserves headline capability on small tests but breaks down on long, reasoning-heavy workloads.
• –For local inference users, this puts real pressure on llama.cpp and other runtimes to add proper sparse-attention support instead of treating V3.2 like a standard dense model.
• –The gap is big enough to matter in practice: 17 points lower accuracy on lineage-512 and 22 points lower on lineage-1024 is not benchmark noise.
• –Because the test was run by the community rather than DeepSeek, it is best read as an important deployment signal, not a final verdict on the model family.
• –The finding also matters beyond Speciale, since the post argues the same limitation could affect other DeepSeek sparse-attention descendants.