NVIDIA Nemotron 3 Diarization is now available on ModelScope—adding live speaker attribution to existing ASR workflows without replacing the transcription model. 🎙️ 🤖 👥 Processes streaming audio and returns speaker labels and timestamps for up to eight speaker slots in a single conversation. ⚡ Its end-to-end streaming architecture avoids separately combining voice activity detection, speaker embeddings, clustering, and post-processing. 🧠 The 99.2M-parameter model uses a 31-layer Transformer encoder with RoPE and builds on NVIDIA’s Streaming Sortformer architecture. 🔌 Pair it with Nemotron ASR, Parakeet, Canary, Whisper, or another ASR system to create speaker-attributed transcripts. 🏢 Designed for meetings, contact centers, clinical conversations, live captioning, media analysis, and multi-party voice agents. 🖥️ Supports NVIDIA Ampere, Hopper, and Blackwell GPUs, with inference through NeMo Speech C++.

Digital PDFs or warped phone photos, TeleOCR parses them with one lightweight 1.2B vision-language model. 📜 Apache 2.0 License. 🤖 📃 🏆 Scores 96.87 overall on OmniDocBench v1.6, the highest among the listed specialized VLMs, and ranks #1# in the ICDAR 2026 Sci-ImageMiner Challenge. 📷 Handles digital, photographed, curved, and degraded documents directly, without a separate dewarping model. 🧠 Combines geometry-aware synthesis, consensus-generated labels, image-based self-verification, and progressive training from vision-language alignment to reinforcement learning. ⚡ Supports structured parsing of text, tables, formulas, layouts, and reading order, with synchronous or asynchronous vLLM inference.

Xiaomi MiMo-V2.6 is now open—a native multimodal agent family built for large-scale reinforcement learning. 🚀📜 MIT License. 🤖 🏆 MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index and reaches 71.9 on DeepSWE v1.1, 89.9 on Terminal-Bench 2.1, and 82.0 on OSWorld-Verified. 🧠 The 1.02T MoE activates 42B parameters and supports text, images, video, audio, and a 1M-token context. ⚙️ One mixed RL run trains coding, general, visual, and cybersecurity agents together. Pro and Flash completed 30 steps each in under six days, producing around 750K trajectories. 🌐 The models support computer use, 3D creation, embodied control, coding, design, video, and music workflows.

Shanghai AI Lab and SJTU’s LUMIA Lab release NCP-ArchPreview, an open-weight 8.9B language modell. 📜 Apache 2.0. 🤖 📄 ⚡ Trained on 5.73T Dolma 3 tokens, it reaches OLMo-3-7B’s final Stage 1 loss with only 51.3% of the tokens, a 1.95× convergence gain. 🏆 Its Stage 1 macro-average rises from 46.59 to 49.04, with +5.99 on GSM8K and +4.28 on HumanEval. 🧠 NCP jointly predicts tokens and a concept sequence at one-quarter the length, then feeds those concepts back to guide generation. 🛠 Domain adaptation updates only the 17M-parameter concept module while keeping the token backbone frozen. 🚀 Concept-conditioned drafting improves mean accepted length by 4.17% with negligible overhead.

inclusionAI open-sources the Ming-Image-0.1-Design family, two complementary 6B models for visual-design workflows.📜 MIT License. 🤖 🤖 🏆 Design ranks #1# among open-weight models on the Artificial Analysis UI/UX Design leaderboard. Layer leads all 12 evaluated Crello settings and runs 4.3× faster than the evaluated 20B open-weight baseline. 🎨 Design generates complete UIs, dashboards, infographics, and posters at up to 2048×2048, with strong text rendering and native transparent RGBA output. 🧩 Layer decomposes flattened graphics into independently editable RGBA layers while preserving the original aspect ratio. 🧠 The pipeline combines multimodal prompt conditioning, a diffusion transformer, and a 4-channel VAE. Layer adds multi-frame generation for separate layer outputs.

Apsara 2026 is ON 🔥 Come find the ModelScope booth, snap a pic, and take home some merch: 🎒 Backpacks · Crossbody bags · Tote bags 🛋️ Neck pillows · ☕ Mugs · 💻 Laptop stands 🧸 Plush pendants · 🧢 Hats · 🧲 Fridge magnets 🎧 Earphone pouches · 📱 Phone chains ...and more. 📍 杭州国际博览中心二期 · 1F · 智能馆 · 1-5C · 魔搭社区 Hangzhou International Expo Center Phase II · 1F · Intelligence Engine · Booth 1-5C · ModelScope Come early before they're gone 👀

Qwen introduces RecreationBench, a benchmark for Hybrid Computer-Use Agents with 250 application-recreation tasks across Ubuntu, macOS, Windows, Android, and Web, spanning domains such as productivity, development, graphics, multimedia, and science. Unlike GUI-only or terminal-only benchmarks, agents must explore a running reference app, recreate it in code, and pass both programmatic tests and VLM-based visual evaluation. The playground is ready. Let’s build! 🚀Dataset:

NetEase-Youdao releases Confucius4-R2T2, a true-streaming ASR model for live captions, simultaneous translation, and voice agents. 🤖 🏆 R2T2 achieves SOTA latency and recognition quality among the evaluated open-source models, while remaining competitive with leading closed-source systems. ⚡ Configurable 80ms–2s chunks deliver 200–600ms average latency with near-offline recognition accuracy. 📝 Append-only decoding commits stable text without revising earlier words, avoiding transcript flicker and giving downstream agents reliable input. 🌍 Optimized for Chinese and English, with multilingual recognition, hotword prompts, and contextual prompts. 🧰 The GitHub repo includes inference code, a minimal example, and a vLLM backend for both offline and real-time streaming ASR. 📜 Code: Apache 2.0. Weights: NetEase Model Use License Agreement.

🎨 Qwen-Image-2.1 is coming, and we're opening 50 early access spots for experienced creators and developers! 🔗 Apply here: 📮 We'll reach out by email if you're in. 💡 Program requirement: publish at least one original showcase or a hands-on review on your social media by Sep 28 at 23:59 (UTC+8). Your honest take, whether glowing or critical, is exactly what helps us make it better.

A 35B co-work agent with only 3B active parameters. Occamy-1.0 is built to carry complex workflows through. 📜 Apache 2.0. 🤖 📄 🏆 Scores 69.10 on AutomationBench, up 29.7 points from Qwen3.6-35B-A3B and ranking #1# in the evaluated 35B-A3B group. 🧭 Maintains context across search, tool calls, terminal coding, files, delegated runs, and history compaction, while retaining strong instruction following. 🧠 Execution-grounded training spans long-horizon interaction, software engineering, and tool-call grounding. Marathon and Sprint experts are merged, then refined with SAO. 🛠️ The open Dressage stack supports multi-harness RL. BF16, FP8, NVFP4, and GGUF checkpoints are also available.

China Telecom open-sources Xing4.0-29B-A4B, a 29B MoE model that activates just 4B parameters for long-horizon agents.📜 Apache 2.0. 🤖 📄 📄 🏆 Leads both comparison models on Claw-Eval, Terminal-Bench 2.1, and DeepresearchBII, while reaching 75.0 on SWE-bench Verified. 🧠 mHC + MLA + MTP supports multi-step planning, tool use, and stable execution across 256K context, expandable to 512K. ⚡ End-to-end training on domestic computing infrastructure, with system optimizations boosting throughput by about 96%. 🛠️ Supports major training, inference, and Agent frameworks, plus cross-architecture deployment and lightweight adaptation for private-domain tasks.

🚀 ZDTaichu5.0-9B is now on ModelScope! 🤖 An on-device multimodal model from TaichuAI. At 9B parameters it runs on a single GPU and brings spatial reasoning, embodied AI and agentic tool use to edge deployment. Qwen3.5-9B backbone + C-RADIOv4-H vision encoder, 128K context, any-resolution image and video input. 🧭 Spatial reasoning: leads the compared 10B-scale open VLMs (Qwen3.5-9B, STEP3-VL-10B, gemma4-8B-E4B) and scores above Gemini 3 Pro, Grok 4 and GPT-5.2 on ViewSpatial, MMSI-Bench and MindCube-tiny 🛠️ Agent: highest among the compared open models on TAU2-Bench, Claw-Eval and IFEval 📄 First-tier results on documents, charts, OCR, visual math and video, with a ready-to-use vLLM branch and Docker image 🧠 Entropy-Gated Adaptive Recurrent Reasoning: extra latent refinement steps go only to the hard tokens

One physical intelligence loop, now in 2B and 8B. PhysBrain 1.5 is here！ 🤖 🤖 📄 🏆 PhysBrain 1.5-8B scores 72.5 across 28 embodied-understanding benchmarks, ranking #1# among the evaluated open-source models and leading 14 tasks. The compact 2B reaches 66.6. 🧭 Both models cover spatial perception, 3D reasoning, embodied planning, grounding, affordance, and trajectory reasoning. 🦾 They generate end-effector trajectories and predict future states as aligned RGB, depth, and robot masks. 🧠 Built on Qwen3-VL, PhysBrain 1.5 models language, spatial outputs, actions, and future states as tokens within one autoregressive backbone. No task-specific heads.

Only ~1.2B parameters active at a time. Edge0-8B-A1B-preview makes an 8B-class MoE practical for local inference.📜 Apache 2.0. 🤖 ⚡ Reaches 23.9–25.3 tokens/s with about 1.0 GiB peak active memory in the reported short-context benchmark. 🏆 Retains most of the FP16 base model’s quality, with an average gap of just 2.8 points across five benchmarks. MMLU-Pro rises from 65.8 to 70.1. 🧠 SSD expert offload keeps most weights outside active memory, while a prerouter predicts which experts each layer will need next. 🛠 Recover-LoRA offsets quality loss from 4-bit quantization and expert offloading. The current Preview runs locally through MLX on Apple Silicon.

TaoMate-H3 turns MiniMax H3 into a low-latency streaming audio-video generator built for continuous creation. 🤖 ⚡ Runs the DiT 11.45× faster and delivers the first playable video 10.60× faster than standard MiniMax H3 in the reported 480×864, 10-second test. 🎬 Generates synchronized video, speech, and sound in five-second chunks instead of waiting for the full sequence. 🔄 Clean KV cache and integrated audio guidance preserve identity, voice, and motion as prompts change across chunks. 🧩 A three-step LoRA supports continuous portrait or landscape generation at 480p, 768p, and aligned 1080p. Inference code and 4-GPU or 8-GPU deployment are included. 📜 MiniMax H3 Community License.

👑 Atria Dawn Preview is here, built to complete real research and engineering work. 📜 MIT. 🤖 ⚙️ Built on a 744B MoE foundation with a 256K context window. Standard and FP8 weights are available. 🔬 Discovery workflows cover evidence gathering, deep research, experiment design, execution, analysis, and recovery from failure. 🏆 Leads the reported comparison on AutomationBench, BFCL v4, CyberGym, DeepSearchQA, and BrowseComp. Scores include 53.8, 77.0, 86.5, 96.0, and 92.5 respectively. 🧩 Creation and delivery capabilities span software, interactive apps, ML systems, visualizations, reports, and presentations. 🛡 Cybersecurity support covers analysis, vulnerability validation, remediation, and retesting in authorized environments.

Marigold V2 turns an image-editing DiT into a single-step model for sharp, detailed dense prediction.📜 Apache 2.0. 🤖 📄 🏆 Best zero-shot results across all evaluated depth datasets among models trained on comparable data. AbsRel improves by 16%–26% over the previous best on KITTI and ETH3D. 🔍 Fur, foliage, fine wires, and object boundaries stay crisp. The two-stage iREPA and SinkLoss recipe tackles the smoothing and flying-pixel artifacts common in diffusion-based depth models. 🧩 The same framework reaches SOTA results in depth completion, see-through depth, surface normals, and intrinsic image decomposition. Depth completion records the lowest RMSE across all four reported benchmarks. ⚡ A pretrained Qwen-Image-Edit DiT becomes a single-step dense predictor through lightweight adaptation. Training takes less than a week on one 32 GB GPU.

Benchmark-leading scientific reasoning meets long-horizon agent capabilities. Intern-S2-397B is now available in BF16 and FP8.📜 Apache 2.0. 🤖 🏆 Scores 87.0 on FrontierScience-Olympiad and 84.0 on SWE-bench Multilingual, leading the reported comparison on both. Several specialized scientific benchmarks also see top results. 🔬 Raw scientific pages become training material directly, preserving text, visuals, symbols, and their relationships without intermediate parsing. 🧪 Joint training across 20+ scientific domains covers scientific reasoning, biomolecular interaction design, material generation, and time-series forecasting. 🛠 Long-horizon agent RL strengthens tool use and sustained execution, with thinking and non-thinking modes available.