Molmo2 Codec

Native visual-space Pruning-16 release: exact training/inference contract, MLVU and six frozen-ID paired benchmarks, checkpoint and training audit.

Visual-first comparison of the learned P-tokenizer, Pruning-16, and Compression-9; includes native-grid mechanics, token efficiency, controlled step100 evidence, and the formal Pruning-16 training contract.

Two-table scorecard: current step500 Letian/LLaVA-OV2 comparison and verified step100 six-task long-video Dense/Codec suite, with checkpoint and live-status separation.

Research-style V6 milestone with real I/P timelines, 15 playable GOP decisions from 31 audited long videos, a Dense-vs-Codec frame-budget audit, Stage-1 reconstruction, formal 8K step-100 performance, and the mixed 32K Stage-2 recipe.

Visual comparison of raw GOPs, effective LLM-visible timelines, P fusion, and token-budget policies; records the final AdaCodec 2FPS/max_p16/16K decision.

Visual audit of predictive-cost I-frame triggers versus Kmax GOP-cap boundaries on 31 long 2FPS IceFrog timelines.

Verified V6-delta Stage-2 step-1200 MLVU record: 144/200 = 72.0% raw accuracy; explicitly separates the 46.9514 aggregate and labels historical step8800 as a different recipe.

V6 evidence snapshot: Stage-1 matched-P reconstruction controls, untuned RGB/Hphi/P boundary probe, mixed-8K Stage-2 regression, and 20-step counterfactual repair with strict P-content gates.

Untuned-LLM MLVU n=200 control: RGB81 61.5%, Hphi81 61.5%, real P16 58.5%, with exact evaluation contract and separation from the earlier native probe.

Unified V6 audit: Molmo-native 144-to-16 P-tokenizer, initialization/freezing evidence, real-video motion oracle results, and explicit Stage-1/Stage-2 readiness gates

Visual audit of 16 direct queries vs 64-to-16 latent topology, block attention mask, Molmo native connector, and the Stage-1/Stage-2 hand-off

Visual audit of the own-search codec front end: ViT-aligned 14x14 blocks, 2-fps +/-48 hierarchical search, adaptive pcost GOP, verified lambda scaling, motion saturation, gamma calibration, and Kmax issues

P0 diagnosis: P-token native-interface mismatch, missing block mask, broken Stage-1/2 bridge; V5 temporal fix is useful but insufficient; V6 pilot recommended before scaling

Root cause = P-tokenizer capacity (0.49M frozen vs 576M); V4 trainable E_P fixes the proxy

Deep architecture comparison: learned delta P-tokens vs codec-guided sparse RGB canvases, with a hybrid recommendation for Molmo2.