MiniMax M3 is now open source! The model combines native multimodal understanding, ultra-long context, and Agent capabilities in one.🚀 New MSA architecture: up to 1M context at 1/20 the per-token compute of the previous gen. 9x faster prefilling, 15x faster decoding, on par with full attention on most tasks. Two versions 👇: MiniMax-M3 (full precision) and MiniMax-M3-MXFP8 (quantized, lower VRAM). 🤖 modelscope.ai/models/MiniMax… 🤖 modelscope.ai/models/MiniMax… 🧠 12hrs autonomous: reproduced an ICLR 2025 Outstanding Paper end to end, 18 commits 23 experiment plots ⚡ 147 iterations, 9.4x CUDA speedup: FP8 matmul kernel on Hopper, peak utilization 7.6% → 71.3%, zero human intervention 🛠️ PostTrainBench: scored 37.1, ranking 3rd behind Opus 4.7 (42.4) and GPT-5.5 (39.3)

DiffusionGemma is here! Google DeepMind’s open-weights 25.2B MoE model brings discrete diffusion to Gemma 4, with 3.8B active params and block-level text generation. 🚀 🤖 modelscope.ai/models/google/… • Speed: parallel denoising generates 15-20 tokens per forward pass, with reported >1100 tokens/s in low-batch FP8 settings • Architecture: encoder caches prompt context, decoder denoises 256-token canvases with bidirectional attention • Context: up to 256K tokens, native system prompt support, function calling, and configurable thinking mode • Multimodal: text image input, video as frames up to 60s, variable image token budgets from 70 to 1120 • Benchmarks: 77.6% MMLU Pro, 73.2% GPQA Diamond, 69.1% AIME 2026 no tools, 69.1% LiveCodeBench v6 • License: Apache 2.0

PP-OCRv6 just landed on ModelScope! PaddleOCR’s new OCR model series scales from 1.5M to 34.5M params and surpasses billion-scale VLMs on OCR tasks. 🚀 🤖 modelscope.ai/collections/Pa… • Models: tiny 1.5M, small 7.7M, medium 34.5M for edge, mobile, and server deployment • Accuracy: medium tier improves over PP-OCRv5_server by 4.6% detection and 5.1% recognition • VLM comparison: surpasses Qwen3-VL-235B and GPT-5.5 on reported OCR benchmarks with only 34.5M params • Languages: one unified model covers Chinese, English, Japanese, and 46 Latin-script languages • Speed: 5.2x CPU end-to-end speedup with OpenVINO, 6.1x on Apple M4 tiny, 0.13s on A100 GPU • License: Apache 2.0

This one is seriously adorable! 💖 Allin-ModelScopeIP-Inspiration-Qwen2512 turns text prompts into inspirational IP visuals Perfect for character concepts, collectible designs, and creative experimentation. 🎨🖌️ 🔗modelscope.ai/models/orangeH… A must-try if you’re exploring new IP styles, toy design, or creative LoRA inspiration. • Format: BF16 safetensors LoRA • Checkpoints: ckpt-5, ckpt-10, ckpt-15, ckpt-20, about 236MB each • Task: text-to-image synthesis on Qwen-Image • License: Apache 2.0

🚀 Multi-Node RL Training with Megatron backend is now live in ms-SWIFT: swift.readthedocs.io/en/late… With our newly introduced Ray extension, orchestrating multi-node RL training on a Ray cluster is now as simple as a familiar YAML config, offering: 🔀 Flexible GPU Allocation | Colocation or separation of training & inference nodes. 🧠 Comprehensive RL Support | GRPO/DAPO/GSPO/SAPO etc., with multi-turn agentic RL supported. ⚡ Megatron Parallelism | Scale seamlessly with TP, PP, CP, and EP. 🎓 GKD / OPSD | Top-k & full-vocab teacher logits, with versatile grouping of GPUs. 🤖 Extensive Model Coverage | 300 LLMs & 200 MLLMs (Full-param & LoRA), covering all frontier models. Learn more at: 💻 GitHub: github.com/modelscope/ms-swi… 📖 Docs: swift.readthedocs.io/en/late… 📝 Examples: github.com/modelscope/ms-swi…

Nex-N2-Pro API inference is now live on ModelScope! 🚀 Try it directly through the ModelScope API and start building with it today! modelscope.ai/models/nex-agi…

Obsessed with this one. 👀 NewChinese_IPfigurine by @WarmBlood_Aban turns character ideas into collectible-style visuals with a fresh New Chinese aesthetic. 🏮 A fun one for character design, toy concepts, and original IP exploration. 🔗 Try it on ModelScope: modelscope.ai/models/zhouwen… • Model size: 2.12GB • Training data: 32 samples • Training steps: 2,000 • Base model: Qwen-Image-2512 LoRA • License: Apache 2.0

New on ModelScope! Meet SCAIL-2, ZhipuAI’s studio-grade character animation model for reference-image driven motion transfer with improved stability, clarity, and native long-video generation. 🚀 🤖 modelscope.ai/models/ZhipuAI… 🎬 3D pose control: preserves depth, orientation, and spatial relationships 📽️ Long video: improved stability and clarity over SCAIL-Preview 🤸 Hard motions: handles flips, turns, occlusion, and multi-character interactions 🧩 Workflow: reference image driving video rendered 3D pose sequence, with ComfyUI support 📜 License: Apache 2.0

OSCAR INT2 KV cache by @ZhongzhuZhou makes ultra-long context more practical on local devices. 🔥🚀 🔗modelscope.ai/models/togethe… 💾 KV memory: reduces Gemma 4 12B-it’s KV cache at 256K context from ~24 GiB to ~3 GiB, saving ~21 GiB 📉 Compression: ~8× smaller KV footprint with q2_0 INT2 KV cache 🧠 Method: calibrated rotations use query covariance for Keys and score-weighted value covariance for Values 🎯 Quality: pushes quantization noise into attention-insensitive directions to preserve near-f16 KV behavior 🖥️ Deployment: Gemma 4 12B now supports both ready-to-run INT2 KV GGUF models through the llama.cpp fork and INT2 KV cache through SGLang. The local path also includes Apple Metal support and fused mixed-precision flash attention. 🧩 Model support: Qwen3 with head dim 128 and Gemma 4 with head dim 512, including Gemma 4 12B-it with sliding-window layers

Nex-N2 is now open source！An agentic model series from Nex AGI built for coding, tool use, deep research, and long-horizon workflows. 🧠🔎 🛠️ modelscope.ai/models/nex-agi… ⚙️ modelscope.ai/models/nex-agi… ● Models: Nex-N2-Pro 397B total, 17B active; Nex-N2-mini 35B total, 3B active ● Agentic Thinking: adaptive reasoning depth coherent reasoning across coding, search, tool calling, and execution ● Efficiency: Nex-N2-mini saves roughly 20% overall token cost vs forced thinking while matching or slightly exceeding task performance ● Open-model lead: 75.3 on Terminal-Bench 2.1, 80.8 on SWE-Bench Verified, 83.7 on BrowseComp, and 1585 on GDPval among listed open baselines ● Deployment: customized SGLang fork, reasoning parser, tool-call parser, Docker image ● License: Apache 2.0

Introducing Claw-Anything by @haiyang73756134 , a benchmark and training set for always-on personal assistants with broader access to the user's digital world. 🚀 🤖 modelscope.ai/datasets/Liber… 📄 modelscope.ai/papers/2605.26… ● Benchmark: 200 human-verified eval tasks across skill, tool, and GUI settings ● Training: 2,015 auto-generated environments across 25 personas, 4 execution dates, and 2 difficulty levels ● Context: long-horizon event streams, interconnected services, and CLI Android GUI interaction ● Gap: GPT-5.5 reaches only 34.5% pass@1, showing current agents still struggle with broad always-on assistance ● Data pipeline: training traces improve the base model by 23.7 pass@1 ● License: Apache 2.0

NAVA just landed on ModelScope！6.3B parameters, one prompt input, synchronized 720p video and stereo audio output. 🔗 modelscope.ai/models/ernie-r… 📝 modelscope.ai/papers/2605.30… 🌟 Benchmarks: SOTA AV sync on VerseBench, with 7.7914 Sync-C and 7.5655 Sync-D, using 2× to 5× fewer parameters than open baselines ⚡ Speed: 720p 1-min generation in about 1 minute via 8-GPU Ulysses sequence parallel🎙️ Audio: dual-channel stereo for scene speech, jointly denoised with video, no post-hoc vocoder alignment 🧩 Control: reference WAVs bind speaker timbre to <S>...<E> spans; camera composition, motion, and pacing follow language prompts 🧠 Resolution: landscape, portrait, and square aspect ratios from the same checkpoint 🧾 License: Apache 2.0

Nemotron 3.5 Content Safety is now live from @NVIDIAAI 🤖One 4B model, three capabilities. 🛡️modelscope.ai/models/nv-comm… ● Multimodal safety: #1 on VLGuard (.903), beating Llama Guard 4 12B (.653) and OpenAI Omni (.216) ● Custom policy: BYO safety rules at inference time, no retraining. Scores .915 on DynaGuardRail safety and .902 on prompt injection ● Multilingual: 12 languages, beating Llama Guard 4 and OpenAI Omni on every multilingual benchmark Open weights and training data are available under the OpenMDW-1.1 license.

Nemotron 3.5 ASR from @NVIDIAAI is built for real-time voice agents. 🌍modelscope.ai/models/nv-comm… Voice agents need to listen in real time. Nemotron 3.5 ASR is built for that. ● Streaming-first with sub-100ms latency, with runtime-configurable latency for live pipelines ● 40 language-locales across Europe, Asia, and the Americas ● Native punctuation and capitalization out of the box Built for voice agents, call centers, meeting transcription, in-car assistants, and live captions.

Meet Nemotron 3 Ultra @NVIDIAAI 's frontier open reasoning model built for long-running AI agents.🚀 🤖 modelscope.ai/models/nv-comm… - 550B total / 55B active parameters, using a hybrid MoE design – only a fraction of the parameters are activated per token. - Up to 1M-token context window. - Up to 5x faster inference and up to 30% lower cost on complex agentic tasks - Open weights, training data and recipes under the OpenMDW-1.1 license. Nemotron 3 Ultra is one to try.

We just added Stable Video Infinity to ModelScope. Infinite length video generation with Error Recycling Fine Tuning, from VITA@EPFL, ICLR 2026 Oral. 🤖 modelscope.ai/models/epfl-vi… 📄 modelscope.ai/papers/2510.09… 🏆Benchmark: SVI-Shot reaches 93.52% subject consistency and 95.86% background consistency on consistent video generation, vs 85.62% subject consistency for FramePack 📽️Ultra-long: at 250s, SVI-Shot reports 97.50% subject consistency with only a 0.63% drop as length scales, while FramePack drops 13.71% 📊Quality: 58.07% aesthetic quality, 62.81% imaging quality, 98.42% motion smoothness 🧵 Control: SVI-Film supports text prompt streams with 1 prompt per 5s clip; SVI-Talk and SVI-Dance add audio and skeleton conditioning 🧠Adaptation: fine-tune lightweight LoRA adapters with limited data to build custom SVI models for new video tasks MIT model repo，Check it out on ModelScope.

JoyAI-Echo is now on ModelScope. Minute-level multi-shot audio-video generation from Joy Future Academy, JD, with paired cross-modal memory for story-level consistency. 🤖 modelscope.ai/models/jd-open… ● Long video: 5 min coherent stories from one prompt JSON ● Speed: 7.5x faster than the original multi-step pipeline with DMD-distilled few-step inference ● Large-scale evaluation: 3,000 generated shots across 100 benchmark stories ● Long-video human eval：preferred over HappyOyster on long-video visual aesthetics 63.6%, audio quality 81.7%, prompt following 80.6%, IP consistency 59.4% ● Short video comparison: preferred over Wan 2.6 on visual aesthetics 58.8% and prompt following 33.8% Under the LTX-2 Community License Agreement，non-commercial use only.

One image, controllable 3D output. TripoSplat from @tripoai a lightweight image-to-3D Gaussian model for fast, controllable 3D asset generation. 🔗modelscope.ai/models/VAST-AI… ✅ Adjustable Gaussian count up to 262,144 ✅ Two core files, ~2,000 LOC total, for lightweight implementation. ✅ Near-zero dependencies, no transformers or diffusers required. ✅ Official ComfyUI support Released under the MIT License for flexible commercial and non-commercial use. For image-to-3D and Gaussian Splatting workflows, TripoSplat offers a practical way to reduce setup cost while keeping quality and rendering cost controllable. 📚Paper：modelscope.ai/papers/2605.16…