The most interesting thing today
upstage/Solar-Open2-250B is trending at #4 with only 362 downloads but 486 likes and a trending score of 415, two days after release, link - likes far outpacing downloads on a 250B model usually means people are bookmarking it to try later, not running it yet. Worth watching whether download counts catch up.
What shipped
- baidu/Unlimited-OCR (image-text-to-text) - 2.4M downloads, 2,920 likes, trending 826, 34 days old. Link
- thinkingmachines/Inkling (image-text-to-text) - 25k downloads, 1,517 likes, trending 586, 9 days old. Link
- poolside/Laguna-S-2.1 (text-generation) - 13k downloads, 537 likes, trending 529, 10 days old. Link, plus a quantized unsloth/Laguna-S-2.1-GGUF already at 29k downloads and 152 likes after 2 days (link).
- zai-org/GLM-5.2 (text-generation) - 596k downloads, 4,388 likes over 37 days, still trending at 287. Link
- Kwaipilot/KAT-Coder-V2.5-Dev - brand new today, 0 downloads but already 78 likes. Link
- microsoft/Mage-Flow (text-to-image) - 411 downloads, 193 likes, 2 days old, with a companion Mage-Flow-Edit-Turbo (image-to-image) at 361 downloads, 63 likes. Model, Edit-Turbo
Datasets: openbmb/UltraX-Preview, 6k downloads, 263 likes, documenting adaptive programmatic editing for pretraining data at scale (link). HuggingFaceCode/stack-v3-train, released 1 day ago, already 64 likes on 0 downloads (link).
What was published
- 3D-Aware VLMs with Implicit and Explicit Geometries - argues 2D-trained VLMs struggle on 3D spatial reasoning tasks and proposes combining implicit and explicit geometric representations to fix it. arXiv
- GraphVid: Interactive Graph-Controllable Video Generation - replaces text/motion prompts with a graph structure to specify precise multi-object interactions in generated video. arXiv
- Barzilai-Borwein Fails Superlinear Convergence on an Open Set of Quadratics for Every $n\geq 4$ - proves the popular BB optimization method does not achieve superlinear convergence on a nontrivial set of quadratic problems in dimensions 4 and up. arXiv
- Expanding Flow Maps - extends flow-based generative models beyond fixed-dimension parameterizations. arXiv
- Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity - argues time-series explanations need to show necessity, not just sufficiency, for a subsequence to preserve a classifier's prediction. arXiv
The connection
The model side is all incremental point-releases (Laguna-S-2.1, GLM-5.2, Mage-Flow) getting quantized and repackaged within days of launch, while the paper side is picking apart foundational assumptions - that 2D VLMs generalize to 3D, that BB converges superlinearly, that sufficiency is enough for explainability. Today's shipping is iteration; today's research is subtraction.
Generated by agentmailkit, locally.