#11
Baidu's **Unlimited-OCR** is an advanced OCR system designed to push the boundaries of document parsing through "one-shot long-horizon parsing," positioning itself as a significant step beyond DeepSeek-OCR. Released in late June 2026 with rapid community adoption, it supports single-image inference in two modes ("gundam" for cropped high-resolution and "base" for full-page), as well as multi-page and PDF document parsing. The model is deployable via HuggingFace Transformers, vLLM, and SGLang backends, with Docker images provided for CUDA 12.9 and 13.0 environments, making it highly flexible for production use. It features an OpenAI-compatible streaming API, batch inference scripts, and n-gram repetition suppression for cleaner outputs. The project has gained rapid traction on GitHub Trending, is available on Baidu Cloud, Hugging Face Spaces, and ModelScope, and supports training via ms-swift. Its notable momentum—community integrations within days of launch, a published arXiv paper, and backing from a major tech company—makes it a strong trending candidate. It builds on PaddleOCR and DeepSeek-OCR lineage, representing Baidu's push into state-of-the-art multimodal document understanding.
ModelsServingAppsmultimodalvisionlocal-first
★ 19,382 total+4,696 this week
#39
LingBot-Map is a feed-forward 3D foundation model developed by the Robbyant Team for streaming 3D scene reconstruction from video sequences. At its core is the Geometric Context Transformer (GCT), a novel architecture that unifies coordinate grounding, dense geometric cues, and long-range drift correction within a single streaming framework. It achieves this through three key mechanisms: anchor context for coordinate grounding, a pose-reference window for dense geometric cues, and trajectory memory for long-range drift correction. The model is designed for high-efficiency streaming inference, leveraging a paged KV cache attention mechanism — powered by FlashInfer — to achieve stable inference at approximately 20 FPS on 518×378 resolution over sequences exceeding 10,000 frames. LingBot-Map claims state-of-the-art reconstruction performance across a wide range of benchmarks, outperforming both existing streaming methods and iterative optimization-based approaches. The project is evaluated on an impressive breadth of datasets including KITTI, Oxford Spires, VBR, Droid-W, TUM-D, 7-Scenes, ETH3D, Tanks and Temples, and NRGBD, all of which have been released as part of the evaluation benchmark. The repository provides multiple model checkpoints: a balanced general-purpose model used in the paper, a long-sequence-optimized variant, and a stage-1 training checkpoint compatible with the VGGT model for bidirectional inference. Installation is straightforward via conda and pip, with PyTorch 2.8.0 and CUDA 12.8 as the recommended stack. The interactive demo uses a browser-based viser viewer and supports features like sky masking via an ONNX segmentation model, keyframe interval control to manage KV cache memory, windowed inference for very long sequences exceeding 3,000 frames, and video input with configurable FPS. A notable demonstration includes a 25,000-frame, 13-minute indoor walkthrough rendered through an offline batch rendering pipeline using NVIDIA Kaolin. The project is actively maintained, with recent updates including bug fixes for SDPA and FlashInfer KV cache issues, a torch.compile acceleration flag, and ongoing plans to release a stronger model supporting even longer sequences. Models are hosted on both HuggingFace and ModelScope, making them accessible to a broad international audience. The project is licensed under Apache-2.0. Its trending appeal likely stems from the combination of real-time streaming 3D reconstruction capability, long-sequence stability, broad benchmark coverage, and practical usability with interactive demos — addressing a highly competitive and commercially relevant problem space in computer vision and robotics.
ModelsAppsServingmultimodalvideovision
★ 15,592 total+1,890 this week
#88
turbovec is a Rust-based approximate nearest neighbor (ANN) vector index with Python bindings, built on Google Research's TurboQuant algorithm — a data-oblivious quantizer that requires no separate training phase and achieves near-optimal distortion. Its headline claim is dramatic memory reduction: a 10-million-document corpus that would consume 31 GB as float32 fits in just 4 GB, while still searching faster than FAISS in most configurations. The core approach normalizes vectors onto a unit hypersphere, applies a random orthogonal rotation to make coordinate distributions predictable (converging to a known Beta/Gaussian distribution), then applies per-coordinate TQ+ calibration (fitting shift and scale scalars from empirical quantiles) before Lloyd-Max scalar quantization at 2-bit or 4-bit widths. Because the rotation makes the distribution data-oblivious, no training corpus is needed — vectors can be added online at any time without rebuilds or parameter retuning. Search is accelerated by hand-written SIMD kernels targeting ARM NEON and x86 AVX-512BW, beating FAISS IndexPQFastScan by 10–19% on ARM and winning 4-bit configurations on x86. A key practical feature is native filtered search: an allowlist of IDs or a slot bitmask is honored directly inside the SIMD kernel at 32-vector block granularity, short-circuiting entire blocks with no allowed slots and avoiding the over-fetching recall penalty common in post-filter approaches. The library exposes both a raw TurboQuantIndex and an IdMapIndex supporting O(1) deletes via stable external uint64 IDs. Drop-in integrations are provided for LangChain, LlamaIndex, Haystack, and Agno, replacing their default in-memory vector stores with no pipeline rewiring. Recall benchmarks on OpenAI embeddings (d=1536, d=3072) show TurboQuant beating FAISS IndexPQ by 0.2–1.9 points at R@1. The project is available on PyPI and crates.io, backed by an arXiv paper, and targets privacy-sensitive or resource
ModelsRetrievalvector-databaselocal-firston-device
★ 14,412 total+810 this week