Nvidia DLSS 5 official model outperforms leaked version by up to 90%

By Splitfeed EditorialLast updated Oct 19, 2026, 07:10 UTC

On October 18, 2026, Nvidia released its official DLSS 5 neural rendering model, and performance figures published by Tom's Hardware showed it outperforming a previously circulating leaked version by up to 90% when used in games without official DLSS 5 support. Nvidia stated that the new model processes neural workloads up to two times faster than its predecessor and reduces VRAM consumption by 20%. A day later, TweakTown reported on mod-based implementations that apply the official model to unsupported titles, with those mods producing the same headline gains.

The context that gives these numbers weight is the modding pipeline itself. DLSS has historically required game-by-game integration, leaving large portions of a player's library outside the feature's reach. The fact that community-built mods can now slot in Nvidia's latest official model, and see gains of this scale over a leaked build, suggests the gap between a finalized model and a pre-release or unofficial one is larger than many observers expected. The 20% VRAM reduction matters separately, because VRAM capacity has been a persistent point of contention between Nvidia and its critics, and any reduction in memory pressure on existing hardware is relevant to users who cannot upgrade.

None of the three camps have published reactions to this story yet. The Pro-AI camp would typically treat results like these as straightforward evidence that AI-driven upscaling and neural rendering are maturing rapidly, with efficiency gains compounding over short release cycles. The Anti-AI camp would likely raise questions about what these models are trained on, whether the gains hold consistently across hardware tiers, and whether dependence on proprietary neural rendering locks players into a single vendor's ecosystem. The Middle Ground camp would probably welcome the performance and VRAM improvements while noting that mod-based delivery of features, rather than official support, places the burden of compatibility and stability on unpaid community developers.

The figures worth tracking are independent benchmark comparisons across a wider range of GPUs and game titles, which would show whether the up-to-90% ceiling reflects typical gains or an optimistic edge case. Official Nvidia documentation on which hardware generations can access the full 2x neural workload improvement would also clarify how broadly these numbers apply.