ModelScope

Intern-Decision scales multimodal structured decision-making across 0.8B, 2B, and 4B models. 🤖 https://modelscope.ai/collections/Shanghai_AI_Laboratory/Intern-Decision 🏆 The three models average 79.38, 84.68, and 90.02 across seven decision benchmarks. Intern-Decision-4B surpasses Jev at 88.74 while achieving better probability calibration. ⚡ Reported mean latency is 33.98 ms, 33.28 ms, and 44.16 ms respectively, compared with 109.70 ms for Jev in the same local HF setup. 🧠 One forward pass answers multiple choice, score, and yes/no questions with typed JSON and calibrated probability distributions. 🖼️ Text states and up to eight images enable routing, tool selection, scoring, and multimodal workflow control. 📜 Apache 2.0. Applicable Qwen3.5 upstream notices must be retained.

Jina-OCR-v1 parses entire pages into structured Markdown at 2.57 pages per second.🚀 🤖 https://modelscope.ai/models/jinaai/jina-ocr-v1 📄 https://modelscope.ai/papers/2609.03181 🏆 Scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, 7.4 points above DeepSeek-OCR on the latter. ⚡ Delivers the highest throughput among 14 evaluated systems at concurrency 32. 📑 Preserves text, formulas, tables, and reading order in a single pass. 🧠 The 3.4B MoE activates only 570M parameters per token, while FastMTP accelerates decoding without changing the autoregressive output. 📜 CC BY-NC 4.0. Commercial use requires permission.

Two LoRAs, one clean layer-editing workflow: extract an object as a transparent layer, then remove it from the original image. 🤖 https://modelscope.ai/models/DiffSynth-Studio/Qwen-Image-2.1-LayerExtract 🤖 https://modelscope.ai/models/DiffSynth-Studio/Qwen-Image-2.1-LayerRemove ✂️ LayerExtract isolates a prompt-specified subject and exports it on a transparent background for independent editing. ✨ LayerRemove takes the source image and a reference layer, removes the corresponding object, and reconstructs the scene behind it. ⚡ Both are compact Qwen-Image-2.1 LoRAs that can be hot-swapped within the same DiffSynth-Studio pipeline. 🧩 Together they turn flattened images into editable assets for compositing, repositioning, replacement, and design workflows. 📜 LoRA weights: Apache 2.0. Qwen-Image-2.1 base-model terms also apply.

IQuest-Q1 lands with 320B MoE scale, 15B active parameters, and a 512K context window for long-horizon coding agents. 🤖 https://modelscope.ai/models/IQuestLab/IQuest-Q1 🧠 Built for agentic coding, complex reasoning, and multi-step tool use across long-running workflows. 🏆 Scores 84.5 on CyberGym, 83.2 on Terminal-Bench 2.1, 64.6 on DeepSWE v1.1, and 63.0 on NL2Repo. ⚡ 256 experts with 8 activated per token, hybrid sliding-window/full attention, and recursive MTP for faster decoding. 🛠️ Production deployment supports SGLang and vLLM, with integrations for Claude Code, Codex CLI, and OpenAI-compatible APIs. 📜 Weights released under the IQuest-Q1 License.

Audio8 ASR Infinite keeps transcribing around the clock with bounded memory and latency. 🤖 https://modelscope.ai/models/Edge0/Audio8-ASR-Infinite 🎙️ A rolling KV Cache enables unlimited-length Chinese and English transcription without accumulated drift. 🏆 At 480 ms delay, it reaches 1.75 CER on AISHELL-1, 2.89 on AISHELL-4, and 3.04/6.81 WER on LibriSpeech test-clean/test-other. ⚡ Select an 80, 120, or 160 ms audio clock and tune delay from 240 to 560 ms. Semantic VAD distinguishes pauses from real turn endings. 📜 Preview release under Apache 2.0, with continuous deployment through an adapted vLLM stack.

A 5B world model built for real-time interaction. Astronex-World 1.0 is now available.📜 Apache 2.0. 🤖 https://modelscope.ai/models/Astronex/Astronex-World 📊 Scores 70.0 on WBench Full, outperforming evaluated models with 13.6B and 14B parameters while coming within 0.9 points of a 22B model. ⚡ Streams 832×480 video at 24 fps in real time on a single 48GB GPU, with 1280×704 generation also supported. 🎮 Camera trajectories, 64-D continuous actions, embodiment IDs, and timed text events provide frame-aligned control over how the world evolves. 🧠 One backbone supports both bidirectional full-context generation and block-causal streaming with cross-block KV caching. 🛠️ The complete five-stage training pipeline runs on two 48GB GPUs; weights, code, sampling presets, and post-training recipes are included.

Having issues with Qwen-Image-2.1? Our official feedback form connects you directly with the Qwen Image research team. Share your prompts, images, or workflows to help us investigate and improve. 📝https://alidocs.dingtalk.com/notable/share/form/v01WgZOZA5DaVQPeqLX_dv19yqvsgs3oebp3pcjys_1qX0QQ0?source=link English & Chinese welcome.

@InSpatio_AI releases InSpatio-World 1.5, turning images and videos into real-time, explorable 4D worlds. 🤖 https://modelscope.ai/models/InSpatio/inspatio-world-v1.5 📄 https://modelscope.ai/papers/2604.07209 🏆 The 1.3B model scores 68.72 on WorldScore-Dynamic, ranking #1 among evaluated real-time and interactive methods while reaching up to 24 FPS. 🌐 A single image, four images, a panorama, or a video can become a navigable scene with wide viewpoint changes and target camera trajectories. 🎥 Next-view prediction preserves motion, spatial structure, and visual style across space and time, with support for bullet-time creation. 📐 For video inputs, Depth Anything 3 automatically estimates missing depth and per-frame camera intrinsics and extrinsics. 🧠 Spatiotemporal autoregression maintains a persistent world state for consistent long-horizon exploration. 📜 Code: Apache 2.0. Dependencies retain their own licenses.

IndexTeam releases Index-Translate, a multilingual translation family for text, speech, dubbing, and long documents. 🤖 https://modelscope.cn/collections/IndexTeam/Index-Translate 🏆 Index-Translate-9B scores 0.8789 on FLORES, 0.8209 on instTrans, and 0.7387 on MEME, showing strong general translation, instruction following, and cultural-context understanding. 🌍 Covers 150 languages while preserving terminology, formatting, structured data, and protected text instead of simply translating everything. 🎙️ Index-Echo supports S2TT and voice-conditioned S2ST; Index-Homura controls translation length for dubbing, with 81.92% of SandGlass samples staying within 10% of the target syllable count. 📚 Index-NativeLong translates complete documents in one pass, keeping names, terminology, and context consistent across books and long transcripts. 📜 2B and 9B models are released under Apache 2.0, with 35B-A3B preview results also available.

Step 5 Preview is going open source! ⏳ A flagship foundation model built for real-world agentic tasks. Join the countdown: https://modelscope.ai/models/stepfun-ai/Step-5-Preview

Spark-X2.6-4B has landed: an open-source model with 1M context, three thinking budgets, and deployment-ready formats for local agents.@SparkLLM 🤖 https://modelscope.ai/collections/XHToken/Spark-X26 🏆 Stronger at 4B: In the team’s reported comparison table, it reaches a 63.3 average, ahead of Spark-X2.5-4B at 52.1 and Qwen3.5-9B at 49.1 across the evaluated suite. 🧠 Set the thinking budget: Switch between none, low, and high thinking modes to balance reasoning quality, latency, and cost for each task. ⚡ Built for efficient local inference: A hybrid attention design combines 1M context with resource-conscious deployment, while FP8, INT8, INT4, and GGUF releases offer practical precision and footprint tradeoffs. 🛠️ One model, multiple workflows: Use the Instruct checkpoint for everyday agent work or the Base model for continued training and customization. Supports Transformers, SGLang, vLLM, llama.cpp, Ollama, MLX, and LM Studio. 📜 Apache 2.0.

Nanbeige4.3-33B is here with 512K context and strong agentic performance in a 33B dense model. 🤖 https://modelscope.ai/models/nanbeige/Nanbeige4.3-33B 🏆 SOTA agentic performance at the 30B scale: Nanbeige4.3-33B leads most benchmarks in the team’s comparison with open models including Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. 💻 Agentic coding: Scores 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE, compared with 73.0 and 42.5 for Qwen3.8-27B. 🧰 Professional workflows: Reaches 54.2 on Automation Bench and 44.8 on JobBench, covering tasks across tools, occupations, and industries. 🧠 Flexible reasoning and tool use: Switch between thinking and non-thinking modes, and retain previous reasoning for multi-turn agent workflows. 🔬 A teacher for smaller agents: Nanbeige4.3-33B is the flagship teacher model for the team’s planned distillation into smaller models. 📜 Apache 2.0. Supports Transformers, SGLang, and vLLM.