New models #994

First round

• PaddleOCR-VL. ocr. New task: no. Layout plus per-region recognizer, cleanest fit to our regions contract. Apache. Ported by @EHxuban11 on 2026-11-05 in Add PaddleOCR-VL document parsing on the ocr task #999.
• PP-DocLayout V3. detect. New task: no. Layout boxes with reading order. Ships with PaddleOCR-VL. Apache. Ported by @EHxuban11 on 2026-11-05 in Add PP-DocLayoutV3 document layout detection with reading order #995.
• DeepSeek-OCR. ocr. New task: no. Grounding mode gives boxes plus text. Markdown mode needs a page-level field. MIT. Ported by @EHxuban11 on 2026-11-05 in Add DeepSeek-OCR and DeepSeek-OCR-2 to the ocr task #996.
• DeepSeek-OCR-2. ocr. New task: no. Same family as DeepSeek-OCR. Apache. Ported by @EHxuban11 on 2026-11-05 in Add DeepSeek-OCR and DeepSeek-OCR-2 to the ocr task #996.
• GLM-OCR. ocr. New task: no. Text first, needs the page-level markdown field. License to check. Ported by @EHxuban11 on 2026-11-05 in Add GLM-OCR document parsing on the ocr task #1005.
• MoGe-3. depth, normal. New task: no. Same plumbing as MoGe-2, also returns intrinsics. Ported by @EHxuban11 on 2026-11-05 in Add MoGe-3 depth and normal family #998.
• Sapiens2. pose, semantic, normal, matte. New task: no. 308 keypoints and 29 human parts are new class sets. Code license must pass. Blocked: licence. Custom Sapiens2 License on code and weights (pass-through, field-of-use bans, audit and termination rights); not permissive.
• Falcon-Perception. detect, segment. New task: no. 0.3B and 0.6B, boxes and masks from a text query. TII license to read. In progress by @omkar-334 in Add Falcon Perception VLM family (detect, segment) #1009.
• Video Depth Anything. depth. New task: no. Easy model, but needs video input. Large weights non-commercial, fine for us. Ported by @EHxuban11 on 2026-11-05 in Add Video Depth Anything (depth over frame sequences) #1004.
• SAM 3.1. segment. New task: no. Our SAM3 pipeline plus multiplexed tracking. Tracking needs video input too. Image path ported (video tracking still open, no permissive implementation yet) by @EHxuban11 on 2026-11-05 in Add SAM 3.1 image segmentation (LibreSAM31) #1008.
• MolmoPoint. point. New task: no. 8B, heavy, pixel-precise pointing. TBD by @omkar-334 (?)
• SAM 3D Objects. mesh. New task: no. Single-image object mesh, sibling of SAM 3D Body. Blocked: licence. Code under the SAM License plus GPL dependencies (bpy, pymeshfix); only a wrapper like SAM 3D Body would be possible.

Later

• Nemotron-Parse. ocr. New task: no. NVIDIA document parser.
• Nemotron-Parse 3.5. ocr. New task: no. NVIDIA document parser.
• dots.ocr. ocr. New task: no. Add one or two of the OCR models below once the first OCR models prove the pattern. Blocked: licence. MIT LICENSE plus a separate dots.ocr licence agreement with use restrictions covering code and weights.
• GOT-OCR 2.0. ocr. New task: no.
• Falcon-OCR. ocr. New task: no. Ported by @EHxuban11 on 2026-11-05 in Add Falcon OCR family on the ocr task #1007.
• MiniCPM-OCR. ocr. New task: no.
• Unlimited-OCR. ocr. New task: no.
• Xiaomi-OCR-0. ocr. New task: no. 0.87B, 96.83 on OmniDocBench v1.6. Apache.
• OvisOCR3. ocr. New task: no. Alibaba AIDC, 0.85B, 96.74 on OmniDocBench v1.6. Apache.
• OXR. ocr. New task: no. Blocked: licence. OXR Open Source License is Apache-2.0 plus a commercial licence required above USD 2M annual revenue.
• IndicOCR. ocr. New task: no.
• ColQwen2.5. embed. New task: no. Visual document retrieval, no VLM family under embed yet.
• ColIdefics3. embed. New task: no. Visual document retrieval.
• Qwen3-VL Embedding. embed. New task: no. Visual document retrieval. Ported by @EHxuban11 on 2026-11-05 in Add Qwen3-VL-Embedding 2B on the embed task #1006.
• Nemotron VL Embedding. embed. New task: no. Visual document retrieval.
• PaliGemma. detect, segment. New task: no. Older grounding VLM, low priority.

European models

• LightOnOCR-3 (0.8B, 1B, 4B). ocr. New task: no. LightOn (FR). Best European OCR, 86.3 on olmOCR-bench. Apache code and weights. Assigned to @EHxuban11.
• docTR. ocr. New task: no. Mindee (FR). DBNet and FAST detection, PARSeq and CRNN recognition, LW-DETR layout, TableCenterNet tables, orientation classifiers. Apache code and weights. Assigned to @EHxuban11.
• PyLaia + Doc-UFCN. ocr. New task: no. Teklia (FR). Handwriting recognition and line segmentation, 15+ checkpoints. MIT and BSD code, MIT weights. Assigned to @EHxuban11.
• jina-ocr-v1. ocr. New task: no. Jina AI (DE, Elastic-owned). DeepSeek-OCR fine-tune, loads in the deepseekocr family. CC-BY-NC weights. Assigned to @EHxuban11.
• Docling layout heron and egret. detect. New task: no. IBM Research Zurich. RT-DETRv2 and D-FINE architectures, second layout option next to PP-DocLayout V3. Apache weights. Assigned to @EHxuban11.
• Holo-3.1, Holo2, Holo4. ground. New task: no. H Company (FR). GUI grounding on the Qwen VL architecture. Apache sizes only: all Holo-3.1, Holo2 4B and 8B, Holo4-35B-A3B (Holo2-30B and Holo4-27B are non-commercial). Assigned to @EHxuban11.
• Ministral 3, Pixtral 12B, Mistral Small 3.2. vlm. New task: no. Mistral AI (FR). Apache code and weights. Also a mistral/ provider prefix in LibreLLM. Assigned to @EHxuban11.
• ColPali, ColQwen2, ColSmol, ModernVBERT. embed. New task: no. Illuin Technology (FR). Visual document retrieval, covers ColQwen2.5 above. MIT. Assigned to @EHxuban11.
• Midnight. embed. New task: no. Kaiko (NL). Pathology ViT-g on the DINOv2 architecture. MIT code and weights. Assigned to @EHxuban11.
• H-optimus-0, H0-mini. embed. New task: no. Bioptimus (FR). Pathology ViTs. Apache code. H0-mini weights are CC-BY-NC-ND behind a manual gate: download from upstream, do not mirror. Assigned to @EHxuban11.
• Phikon, Phikon-v2. embed. New task: no. Owkin (FR). Pathology ViTs on the DINOv2 architecture. Apache code. Owkin non-commercial weights only allow sharing unmodified copies: download from upstream. Assigned to @EHxuban11.
• Franca, NAF. embed. New task: no. valeo.ai (FR). Fully open DINOv2-style backbone and a feature upsampler. Apache code. Franca weights are research-only RAIL, redistribution allowed for non-commercial research. Assigned to @EHxuban11.
• Iris-3B. generate, depth, restore. New task: no. Speridlabs (ES). One backbone, three tasks. Apache code and weights. Assigned to @EHxuban11.
• Idol-Turbo. generate3d. New task: no. Speridlabs (ES). New weights for the Idol family. Apache. Assigned to @EHxuban11.
• FLUX.2 klein extra sizes. generate. New task: no. Black Forest Labs (DE). base 4B (Apache), 9B, base 9B and KV (FLUX non-commercial, sharing allowed), small decoder. Aliases on the existing family. Assigned to @EHxuban11.
• FLUX 3 Image Klein + prompt upsampler. generate. New task: no. Black Forest Labs (DE). 5.6B, text to image, editing, layout from boxes. Apache code. FLUX Kommunity weights are gated and unread: read the terms before mirroring. Assigned to @EHxuban11.
• FLUX 3 Action. act. New task: no. Black Forest Labs (DE). 7B robot policy, already in LeRobot. Apache code, same gated FLUX Kommunity weights. Assigned to @EHxuban11.
• REKI-2B. act. New task: no. Kyutai (FR). Robot policy with LIBERO, RoboCasa, RoboTwin fine-tunes. MIT code, CC-BY-NC-SA weights. Assigned to @EHxuban11.
• PAFUSE. pose. New task: no. valeo.ai (FR). 3D whole-body pose lifting. Apache code, weights licence to check. Assigned to @EHxuban11.
• Anny. mesh. New task: no. NAVER LABS Europe (FR). Parametric human body model, SMPL alternative. Apache code, CC0 assets. Assigned to @EHxuban11.
• P-SuperPoint. point. New task: no. Pruna AI (DE). Pruned SuperPoint keypoints. Apache code and weights. Assigned to @EHxuban11.
• P-H3-Image. generate. New task: no. Pruna AI (DE). Image generation built on MiniMax-H3. Blocked: licence. MiniMax H3 Community License: its territory excludes the EU, UK, Korea and the US, plus a revenue cap and a mandatory "MiniMax H3" label.
• SESR. restore. New task: no. Arm (UK). Tiny super-resolution, same niche as QuickSRNet. Apache code, BSD weights. Assigned to @EHxuban11.
• NSFW image detector. classify. New task: no. Freepik (ES). EVA02 backbone. Apache code, MIT weights. Assigned to @EHxuban11.
• FAI-DETR, BisenetFormer, RTMO. detect, semantic, pose. New task: no. Focoos AI (IT). Real-time models. MIT code, weights licence not published: ask them first. Assigned to @EHxuban11.

European models, new task needed

• WaffleIron, ScaLR, Alpine, RangeViT. pointseg. New task: yes. valeo.ai (FR). Lidar point segmentation. Apache code with weights. Assigned to @EHxuban11.
• QueryOcc. occupancy. New task: yes. Zenseact (SE). Multi-camera 3D occupancy, CVPR 2026. Apache code and weights. Assigned to @EHxuban11.
• RoMa, DeDoDe, ColabSfM. match. New task: yes. Linköping University and Ericsson (SE). Image matching. MIT. Assigned to @EHxuban11.
• GENIE (35M to 210M). worldmodel. New task: yes. 1X (NO). Video plus actions in, future frames out. Apache code and weights. Assigned to @EHxuban11.
• OVIE. novelview. New task: yes. Kyutai (FR). New views from one image, built on MoGe-2. README says MIT but the repo has no LICENSE file yet. Assigned to @EHxuban11.
• CoralBay. medical3d. New task: yes. Kaiko (NL). 3D CT foundation model. MIT code and weights. Assigned to @EHxuban11.