OvisOCR2 just released on ModelScope, an exclusive 0.8B end-to-end OCR model for page-level document parsing. 🚀 License: Apache 2.0. 🤖 🏆 OmniDocBench v1.6 SOTA: 96.58 overall, the first end-to-end model to rank #1# on a leaderboard previously led by pipeline systems. 📄 PureDocBench leader: highest Avg3 score at 75.06, showing strong page-level document parsing beyond a single benchmark. 🧩 Structured output: generates natural-reading-order Markdown, formats formulas as LaTeX, tables as HTML, and visual regions as image tags with bbox coordinates. ⚙️ Compact deployment: post-trained from Qwen3.5-0.8B with SFT, RL, and OPD, with vLLM inference support.

Huge congratulations to @MiniMax_AI on the $2B financing milestone! 🚀 MiniMax’s CEO has committed 1% of the company’s total equity to supporting the open-source community over the long term. We look forward to our continued collaboration and to bringing more great models to the open-source community together!🤖

Meet Hy3 on ModelScope! 295B total / 21B active MoE, built for agentic workflows with 256K context and an FP8 variant ready for deployment. 🚀 License: Apache 2.0. 🤖 🏆 Human eval: 270 experts compared outputs in blind real-world workflow tests. Hy3 scores 2.67/4, ahead of GLM-5.1 at 2.51/4, with the clearest gains in frontend development, CI/CD, and data & storage. 🛠️ Agent reliability: stronger tool calling, formatting, and error recovery, with SWE-Bench Verified variance within 4% across major scaffolds. 🧠 Multi-turn gains: issue rate drops from 17.4% to 7.9%, and MRCR rises from 42.9% to 75.1%. ⚙️ Model lineup: Hy3 BF16 instruct model plus Hy3-FP8 quantized instruct model, with vLLM and SGLang deployment support.

DeepSeek-V4-Pro-DSpark lands on ModelScope~ Same DeepSeek-V4-Pro checkpoint, now with an added speculative decoding module for inference experiments. 🚀License: MIT 🤖 📄 🏆 Pro-Max results: 93.5 on LiveCodeBench, 3206 Codeforces rating, 80.6 on SWE Verified, and 83.5 on MRCR 1M 📏 Long-context efficiency: at 1M context, DeepSeek-V4-Pro uses only 27% single-token inference FLOPs and 10% KV cache vs DeepSeek-V3.2 🧠 Architecture upgrades: hybrid CSA + HCA attention for 1M-token efficiency, mHC for stronger signal propagation, and Muon optimizer for faster, more stable training

A note from the ModelScope team 🙏 We’ve read your feedback on the recent API inference campaign. The onboarding and setup process has been bumpy for new users, and the rate limits haven't made things any easier. We’re apologize for this less-than-ideal experience. To smooth things out we’ve published a clearer registration guide and are working to scale capacity for future traffic peaks. 📄 Registration & Setup Guide: 💻 API Inference Docs: More good news is coming: ✨ Free API inference is regularly provided (although with daily limit quota) on ModelScope for trending models. You can find the list of supported models through the guide. 🎯 In the furture, we’ll run more "unlimited API inference" campaigns from time to time, please keep an eye out for our next campaign. Thank you for your patience and feedback. Stay tuned! 👀

New open-source SOTA on agentic coding! 🚀 Ornith-1.0-397B achieves 82.4 on SWE-bench Verified  and 77.5 on Terminal-Bench 2.1, topping every open model in its class and beating Claude Opus 4.7 on both. 🤖 📦 Four sizes (9B to 397B-MoE), post-trained on Gemma 4 / Qwen 3.5, MIT licensed and globally accessible. ✨ Notably, Ornith uses RL to generate not just solution rollouts but also the scaffold that drives them. By jointly optimizing both, the model discovers better search trajectories and produces higher-quality solutions. ⚙️ Deployable on a single 8×80GB node, with vLLM and SGLang recipes in the model card.

Rapidata SVG Benchmark just landed on ModelScope, comparing 30 frontier LLMs on static SVG generation from text prompts, with 1.35M+ human votes across preference, coherence, and prompt alignment. 🚀 🤖 📊 Scale: 188,754 head-to-head comparisons, 500 prompts, 14,872 rasterized SVG images, and 1,355,161 human responses 🎨 Evaluation target: raw SVG markup generated by LLMs, rendered to 768x768 PNGs, then ranked by humans instead of automated metrics 🏆 Overall ranking: Claude Fable 5 Thinking leads with 1232.9 ELO, followed by Claude Fable 5 and Gemini 3.1 Pro Preview License: CC-BY-4.0 for the benchmark prompts, with generated outputs governed by each model provider's terms.

SkyJM-Gen-9B is now live on ModelScope! A 9B visual reward model for text-to-image generation that ranks candidate images with prompt-conditioned rubrics, not a single flat score.🚀 🤖 🏆 Text-to-image judging: tops listed reward models on MMRB2, GenAI-Bench, and GenAI-Bench-Verified, with 72.0 / 74.1 / 84.5 🧩 RubricRM workflow: generates evaluation dimensions, weights, and scoring descriptors for each prompt, then scores both candidate images by dimension ⚙️ Training recipe: rubric-trajectory SFT + dimension-level GRPO for more structured preference judgment Apache 2.0. vLLM and Transformers ready.

Qwen-AgentWorld just dropped two releases on ModelScope! An open 35B total / 3B active MoE world model with 256K context, plus a 7-domain benchmark grounded in real environment observations. 🚀 🔗 Qwen-AgentWorld-35B-A3B 🌍 One model for 7 agent environments: MCP, Search, Terminal, SWE, Web, OS, and Android 🧪 47.73 → 56.39 on AgentWorldBench, surpassing Claude Sonnet 4.6 at 56.04 🧠 Three-stage training: CPT injects environment knowledge, SFT activates next-state prediction reasoning, and RL sharpens simulation fidelity AgentWorldBench 🛠️ Covers 7 domains with 2,170 samples and 22.8 average turns 🔎 Scores predictions on format, factuality, consistency, realism, and quality

Unlimited-OCR is now on ModelScope! A 3.3B multilingual OCR model for one-shot parsing across single images, multi-page documents, and PDFs. License: MIT🚀 🤖 📄 Full-document parsing instead of cropped-region OCR 📏 32K output length for long OCR sequences 🧩 Base and gundam image modes for different document layouts ⚙️ Transformers inference + SGLang serving with OpenAI-compatible streaming requests Built to push DeepSeek-OCR-style document parsing further.

Meet Boogu-Image-0.1 from @BooguTeam , an open-source unified image generation and editing model series from Boogu team. Apache 2.0. 🚀 🤖 🖼️ Photorealistic generation with natural lighting and consistent spatial relationships 📝 Bilingual text rendering (Chinese/English) across posters, UI, brand guidelines, handwriting boards 🎨 Stylized generation: miniature 3D scenes, anime portraits, fantasy visuals, mythological art ✏️ Fine-grained text editing: replace, add, or delete characters with font/color/layout control Research preview. Trained on roughly 10x less data than comparable closed-source systems. Three variants: • Base (dense text rendering, posters, documents) • Turbo (fast generation + photorealism) • Edit (object insertion, replacement, removal, style transfer)

👀 JoyAI-VL-Interaction-Preview just landed on ModelScope! An open 8B model for real-time video-language interaction. License: Apache 2.0🚀 👉 Try it now: 📄 Paper: ✨ Real-time presence: built for live video scenarios where the right answer has to arrive at the right moment, not after a user prompt 🚨 Strongest zone: wins 100% of human pairwise comparisons on monitoring and alerting against both Doubao and Gemini video-call assistants 🧠 Interaction training: trained on 4M+ time-aligned clips labeled second by second for speak, stay silent, or delegate 🛠️ Open stack: releases the 8B model, training recipe, data, and deployable system for building always-present visual assistants

🚀 Build with GLM-5.2 and DeepSeek-V4-Pro for FREE! We're giving developers 8 hours of FREE API inference on ModelScope. 👉GLM-5.2: 👉DeepSeek-V4-Pro:

EBench is here. An indoor VLA manipulation benchmark from InternRobotics for long-horizon, dexterous, and mobile manipulation in Isaac Sim. 🚀 🤖 🎯 Tasks: 26 evaluation tasks across long-horizon manipulation, simple pick-and-place, and teleop-style dexterous tasks 🏢 Dataset: LeRobot-format trajectories with multi-view videos from head, left hand, and right hand cameras 🛠️ Setup: client-server architecture keeps Isaac Sim on the server side while models connect through a lightweight client 📋 Assets: benchmark assets and dataset are released for reproducible embodied evaluation 💻 License: MIT

Quasar-Preview is here. SILX AI's first public Quasar checkpoint is an ~18B sparse MoE with a ~2B active path, built for long-context and decentralized distillation research. 🚀 🤖 • Architecture: Quasar Long Hybrid Transformer with sparse MoE routing, Quasar / Raven / GLA hybrid layers, and a single-loop scaffold • Context: experimental 5M-token config with Safe NoPE / DrOPE-style staging; long-context path has <1B extension tokens so far • Training: >1T and <1.5T tokens, designed as the starting checkpoint for Bittensor SN24 decentralized training • Benchmarks: 68.4% MMLU 5-shot, 71.4% MATH-500 4-shot, 81.9% PIQA, 74.0% HellaSwag • Caveat: early preview checkpoint, not a final production or SOTA model • License: MIT

LoopCoder-V2 is here, a 7B PLT coding model that reuses shared Transformer blocks for one extra round of test-time reasoning. License: Apache 2.0. 👉 Try it now: ✨ 2-loop sweet spot: stronger than the 1-loop baseline across coding, software engineering, terminal, and tool-use benchmarks 🛠️ Repo-level gains: biggest improvements show up on SWE-bench Verified and Multi-SWE 🌍 Code-first training: trained from scratch on 18T tokens with 100+ programming languages

🎹 Diffusion magic is taking over the music world! 🚀 With Diffusion-Studio, we’ve just launched a powerful two-way music workflow: ✨ Vocals to Music: Generate high-quality backing tracks from clean a cappella vocals. ✨ Music to Vocals: Extract crisp vocal tracks from any full mix. 🔄 The Ultimate Remix Loop: Use Demucs to isolate the vocals, then let vocals2music craft a brand-new accompaniment inside DiffSynth-Studio! 👉 Try it out now: (💡 Pro-tip: Works best on vocals with a stable beat and strong rhythm!)

Twinkle is now at v0.4.0! 🔥 The fully open-sourced solution for multi-tenant Training-as-a-Service, with Tinker API compatibility. Now packed with broader model coverage, more training algorithm support, and an improved backend built to scale. Here’s what’s cooking: 🐳 DeepSeek V4 Support: Flash FSDP2 + Expert Parallelism (EP) training, plus native tool-call parsing and cleanup. 🤖 Qwen3.5 Evolution: Maximize efficiency with padding-free / packed-sequence support and MoE GatedDeltaNet sequence parallelism. 🔮 Gemma 4: Full multimodal training support is officially here, complete with a fresh 12B cookbook! 🧬 LoRA Level-up: Added rsLoRA for Multi-LoRA, FSDP2 for Multi-LoRA SFT, and EP LoRA SFT examples for DeepSeek V4 and Qwen3.5 MoE. ⚡ NPU Acceleration: Huge stability and speed gains with fused operators (RMSNorm, RoPE, SwiGLU, SDPA) and FLA patches. Time to supercharge your cluster and squeeze out every ounce of compute. 🏎️💨 👉 Check out the full release notes at and drop us a ⭐ on GitHub: ❤️

New on ModelScope! Anima Tile & Repair ControlNet-LLLite v1.0, a small ControlNet for anime image repair that preserves the original composition and character structure. 🚀 🤖 🎨 Fixes blur, compression artifacts, noise, and low-quality degradation ✨ Tile-based local detail enhancement and sharpening 🖼️ Keeps original composition and character structure intact during repair ⚡ Python and ComfyUI ready V2 is in training with expanded task coverage, coming next week.