NVIDIA's LocateAnything replaces token-by-token box decoding with parallel decoding, hitting 12.7 boxes/sec on one H100

thisguyknowsai · x · 2026-09-10

• The bottleneck: Today's VLMs decode bounding boxes as 4 sequential coordinate tokens (x1, y1, x2, y2), tearing box geometry into 1D tokens that don't understand each other — the real reason grounding models are slow.
• The fix: NVIDIA's LocateAnything introduces Parallel Box Decoding, treating the whole box as an atomic unit and predicting all 4 coordinates in a single forward pass.
• Architecture: Built on a Moon-ViT vision encoder + Qwen2.5 decoder, with a Fast Mode for speed and an automatic Slow Mode fallback that triggers on format irregularities or spatial ambiguity.
• Performance: 12.7 boxes/sec on a single H100.

Original post →

More from Models

• Sol61 flunks instruction-following: skips mandated research even when explicitly told — lucasmeijer · 2026-10-26
• Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
• Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
• Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
• supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
• Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26

Models channel →