ModelScope @ModelScope2022

@deepseek_ai Huge congrats on the open release!

DeepSeek-V4.1-Flash-1017 delivers stronger coding and agent performance with faster task completion and shorter outputs. Available on ModelScope: 🤖 modelscope.ai/models/deepsee…

DeepSeek @deepseek_ai

22 hours ago

🚀 DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. 🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy,

keyboard_arrow_left Previous keyboard_arrow_right Next

1K 4K 33K 5.5M 6K

NVIDIA Nemotron 3.5 Bolt has landed on ModelScope: a 74B-A9B open-weight VLM in a 24GB footprint. 🤖 modelscope.ai/collections/nv… 🏆 Strong agent results: The 24GB model reaches 50.98 on Terminal-Bench 2.1 and 66.75 on BrowseComp. NVIDIA positions it as the #1 open model in its class on PinchBench and #2 among open models on BrowseComp. ⚡ Built for local inference: Native 2-bit quantization fits the model into 24GB while retaining 9B active parameters per token. 👁️ Multimodal, long-context agents: Accepts text, images, and video with up to 256K shared context for prompts, tool results, and conversation history. 🧩 Choose the right variant: Use 24GB IQ2 for local inference, BF16 for post-training and custom quantization, or LiLiCorr variants for speculative decoding. 📜 OpenMDW License Agreement v1.1. Commercial and non-commercial use supported.

Introducing Hy-Memory-30B-A3B, a dedicated long-term memory writer with around 3B active parameters. 🤖 modelscope.ai/models/tencent… 📄 modelscope.ai/papers/2610.22… 🏆 Cross-framework leader: Scores 55.49 across Mem0, Hy-Memory, MemOS, and held-out EverMemOS in the fixed 20% evaluation matrix, ahead of Claude Opus 5 at 54.62. 🧠 Write memories, not chat replies: Decides what to retain, turns evidence into precise memory entries, and updates the persistent store when new information arrives. 🔄 Transfers beyond its training stack: Leads the held-out EverMemOS harness at 56.89, showing the writer can plug into unseen memory frameworks. 🛠️ Easy to integrate: Runs with vLLM and an OpenAI-compatible API, with a no-thinking mode for structured fact and preference extraction. 📜 Apache 2.0.

Meet MOSS-Image-Base: an 8.56B open-weight text-to-image foundation model built for more controllable generation. 🤖 modelscope.ai/collections/op… 🧩 More than natural-language prompts: Generate from descriptive prompts, tag sequences, or JSON layouts that specify text and object placement. 📝 Layout-aware images: Control captions, scene elements, and bounding boxes for posters, whiteboards, diagrams, and other structured visuals. 🔧 Built for customization: As the base checkpoint in the MOSS-Image family, it is designed for research, fine-tuning, and workflow-specific adaptation. 🖥️ Ready for local workflows: Includes Diffusers support, a Gradio interface, CPU offloading options, and a native ComfyUI plugin. 📜 Model weights and inference code are released under Apache 2.0.

Full GLM-5.3, compressed for multi-node local serving. Kindling AI releases an exlr8 K2/K3/K4 quantization of the 744B MoE model. 🤖 modelscope.ai/models/kindlin… 📦 Flexible memory targets: Choose profiles ranging from 219GB to 378GB, depending on the quality and memory budget of the deployment. ⚙️ One weight set, multiple precisions: K2, K3, and K4 expert weights ship together, letting users change the precision mix at load time without re-encoding the model. 🖥️ Built for distributed local inference: Designed for tensor parallelism across 3–7 unified-memory nodes, with TP=3 and TP=4 configurations already tested. 🧪 Quality-aware allocation: Hessian-calibrated manifests spend additional bits where they reduce divergence from the original BF16 model most effectively. 📌 Preview release. Performance and full BF16-alignment validation are still in progress. GLM-5.3 License applies.

keyboard_arrow_left Previous keyboard_arrow_right Next

NVIDIA Nemotron 3.5 Super VL is open released in BF16 and NVFP4: a 120B MoE multimodal agent model that activates just 12B parameters per token. ModelScope’s ms-swift supports both inference and fine-tuning, helping developers put the model to work and adapt it to specialized multimodal applications. 🤖 modelscope.ai/collections/nv… 💻 github.com/modelscope/ms-… 🏆 Agent performance: Outperforms the 5x larger Nemotron 3 Ultra on the reported Hermes Bench and LangChain evaluations. ⚡ Faster workflows: Delivers up to 5x higher throughput in its size class and completes evaluated agent tasks up to 4x faster. 👁️ Multimodal reasoning: Understands text, images, charts, documents, and video while reasoning, coding, and using tools. 🧰 Built to customize: Fine-tune with your own multimodal datasets using LoRA in ms-swift, or use NVIDIA NeMo for SFT and RL to adapt the model to domain data, tools, policies, and workflows. 🧠 Efficient deployment: NVFP4 with QAD, multi-token prediction, and cross-turn context reuse reduce repeated inference work. Supports vLLM, SGLang, TRT-LLM, and NIM across local, data center, and cloud environments.

keyboard_arrow_left Previous keyboard_arrow_right Next

🌍 Meet BlenderWorld from the Qwen Team, now on ModelScope! Explore 👉 modelscope.ai/datasets/Qwen/… Can visual agents turn real-world references into editable 3D scenes? 🧊 223 Blender tasks across 9 categories 🖼 Text, image & video inputs 📏 2,651 task-specific evaluation rubrics

keyboard_arrow_left Previous keyboard_arrow_right Next

Train your own few-step diffusion models faster, with OPD training workflow now offered in DiffSynth-Studio! Try example “TurTurbo” Qwen-Image-2.1 LoRAs trained with this pipeline: 🤖 modelscope.ai/models/DiffSyn… 📚 Docs: diffsynth-studio-doc.readthedocs.io/en/latest/Trai… ⚡ Fast training feedback: with OPD, we are seeing training convergence typically within 2 GPU-hours. 🎯 Fewer sampling steps: distill FLUX.2-klein-base-4B and Qwen-Image-2.1 from 50 to 8 steps, or Qwen-Image-2.1-Turbo from 8 to 3, while upholding generation qualtiy. 🧠 Simple workflow: train on student-generated trajectories with teacher guidance, without pre-generating teacher trajectories or designing an additional discriminator. 📜 Release under Apache 2.0 license.

keyboard_arrow_left Previous keyboard_arrow_right Next

TaoMate-H3-FL2VA brings 3-step streaming audio-video generation to MiniMax H3. The TaoMate-H3 runtime reports 11.45× faster DiT inference and 10.6× faster time to first playable video than the original MiniMax H3 pipeline in its 480p benchmark. 🤖 modelscope.ai/models/TaoLive… 📄 modelscope.ai/papers/2607.24… ⚡ 11.45× faster DiT: Generates a 10-second 480p sample in 14.8 seconds of DiT time, compared with 169.6 seconds for the original MiniMax H3 pipeline in the reported setup. ▶️ Earlier playback: Produces the first playable video in about 17.3 seconds, a 10.6× improvement over the baseline. 🔊 Synchronized by design: Generates speech, sound, and video together on one timeline instead of assembling them afterward. 🎬 Continuous generation: Processes prompts in five-second blocks while preserving visual identity, voice, and motion across prompt changes. Supports portrait and landscape output at 480p, 768p, and aligned 1080p. 📜 A three-step LoRA for the MiniMax H3 FL2VA base model. MiniMax H3 Community License Agreement applies.

Fish Audio Drama 3 brings voice cloning, multi-character scenes, and in-context audio editing to a 35B-A3B open-weight MoE model. 🤖 modelscope.ai/models/fishaud… 🎭 Point-level control: Place cues such as [sighs], [laughs], and [whisper] exactly where the expression should happen. ✂️ Edit without starting over: Change a word or line while preserving the surrounding voice, timing, and delivery, then continue revising from previous takes. 👥 Built for full scenes: Generate scripts with up to 16 characters, clone voices across languages, and stream long-form speech automatically. ⚡ Production-ready streaming: Delivers the first audio in under 100 ms in the reported single-stream setup and supports high-concurrency serving through SGLang-Omni. 📜 Fish Audio Community License.

Step-5-Preview is now open weight, bringing a 1M-token context window to real-world agentic work. Its 600B sparse MoE architecture activates 27B parameters per token. 🤖 modelscope.ai/models/stepfun… 🏆 Strong open-model coding: Scores 49.0% on StepCodeBench, leading the open-weight models in the reported comparison across code generation, bug fixing, refactoring, and performance tuning. 📈 Built for professional work: Delivers strong results across software engineering, finance, and knowledge tasks, including 66.4 on FrontierFinance. 🧰 From research to deliverables: Coordinates large-scale web research, analyzes complex datasets, and produces finished reports, spreadsheets, websites, and other working artifacts. 🧠 Long-horizon agents: Supports a 1M-token context window, visual input, and workflows extending across thousands of interactions. 📜 Released under the StepFun Community License.

keyboard_arrow_left Previous keyboard_arrow_right Next

Ling-3.1-flash is now available with open weights: a 560B MoE agent model activating around 25B parameters per token, with support for up to 1M context. 🤖 modelscope.ai/models/inclusi… 🏆 Strong agent benchmarks: Leads the reported HealthBench Professional and FrontierSWE comparisons, and ranks second on GDPval-AA v2.1 among the models shown. ⚡ Fast and capable: Scores 41 on the Artificial Analysis Intelligence Index while generating around 210 tokens/s, placing it on the reported intelligence–speed Pareto frontier. 🎨 From prompts to working outputs: Ranks among the top open-weight models in Design Arena for data visualization, React Native, 3D, game development, SVG, and websites. 📦 Easier to deploy: MXFP4 mixed precision reduces the model footprint from around 534GB to 292GB and cuts the tested server setup from 8 accelerators to 4. 🧠 Built for long-running agents: Maintains goals, tool history, and intermediate results across hours-long workflows, with stronger JSON Schema tool calling for production integration. The hosted API currently supports 256K context; the open model supports configurations up to 1M.

keyboard_arrow_left Previous keyboard_arrow_right Next

Introducing EntroPack: a PyTorch tensor compression library that quantizes your model at flexible birrates to fit your exact GPU memory budget. 💻 Open Sourced at: github.com/modelscope/ent… 🎛️ Flexible compression: entropy-coding-based calibration-free quantization that offers arbitrary bpp (bits per parameter), not bound by fixed precisions like 4-bit or 8-bit. ⚡ GPU-native decoding: fast, fused decoding and reconstruction in a single kernel, with minimum overhead, suitable for compute-intensivie workloads such as Diffusion denoising and Transformers-Prefill. 🏆 Make the best use of your GPU VRAM: customized bpp for your exact GPU budget. For example, on a 16G GPU, Qwen-Image-2.1 can be quantized with at 4.8 bpp, resulting in 15.86 GiB usage while cutting reconstruction error from NF4's 9.40% to 4.26%. 📦 Ready to try: pre-quantized weights for Z-Image-Turbo, Qwen-Image-2.1, and MiniMax-H3 at 4 to 8 bpp, for DiffSynth-Studio and ComfyUI. The extreme Z-Image build reaches 2.3 bpp, shrinking its DiT weights to 1.6 GB, around one-seventh of BF16. 📜 Licensed under Apache 2.0. 📄 Technical report: modelscope.ai/papers/2609.34… 🤖 modelscope.ai/models/DiffSyn…

keyboard_arrow_left Previous keyboard_arrow_right Next

JEV-35B turns Qwen3.5-35B-A3B into a fast System 1 decision model, activating only around 3B of its 35B parameters per token. Apache 2.0. 🤖 modelscope.ai/collections/Au… 🏆 Outperforms JEV-27B-VL on the Decision Index public suite while retaining competitive visual decision performance. ⚡ Compare up to 256 text or image options in one pass, return calibrated probabilities, and reach around 204ms median single-request latency in the tested setup. 📦 The NVFP4 version is only 25.6GB and uses around 23GiB for weights, enabling local deployment on a single 32–48GB GPU. 🔍 Quantization preserves 96.6% of the BF16 model’s decisions with nearly unchanged latency and only a small benchmark drop. 🛠️ Start with the included vLLM script. Use /v1/decide for fast System 1 decisions or OpenAI-compatible endpoints for System 2 reasoning.

keyboard_arrow_left Previous keyboard_arrow_right Next

MiniCPM-V 4.7 is now open source on ModelScope: a 35B-A3B MoE multimodal model built to run on-device without compromising performance. Apache 2.0. 🤖 modelscope.ai/collections/Op… 📱 First open-source MoE with smooth, lossless on-device inference, running at 13+ tokens/s on iPhone 17 Pro Max. 🔥 New Pareto frontier: in Instruct mode, it reaches a higher overall score in less inference time than Qwen3.6 35B-A3B. ⚡ Mixed 4x/16x visual token compression lets developers balance visual detail and inference cost per task. 🛠️ Two sizes (35B-A3B and 1B) with FP8, GPTQ, AWQ, BNB, and GGUF variants. Works with SGLang, vLLM, llama.cpp, Ollama, SWIFT, and LLaMA-Factory. 💻 Runs on iOS, Android, and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD, and Arm.

keyboard_arrow_left Previous keyboard_arrow_right Next

DMAD is now open source: 1 to 4 step distillation for image, video, and joint audio-video generation. 🤖 modelscope.ai/models/Zhengmi… 📄 modelscope.ai/papers/2610.02… 🎬 A 1.4 GB rank-128 LoRA turns the 50-step, 33B MiniMax-H3 into a 4-step generator of 1344x768 video with stereo audio, no CFG. 🪶 Consumer GPU ready: 15 s clips on 16 GB, 5 s clips on 8 GB. 64 GB system RAM recommended. 🏆 1.04 FID on 1-step ImageNet-64, 14.47 FID on 4-step SDXL, 85.15 VBench on 4-step Wan2.1-14B, 84.6% human preference over rCM on MiniMax-H3. 🧠 Reframes distribution matching as classification with two discriminator heads, no auxiliary student-score model. 🛠️ Weights for MiniMax-H3, Wan2.1, SDXL, and ImageNet-64, plus training code, H3 training data, Diffusers support, low-VRAM inference, and ready-to-run ComfyUI workflows. 📜 Weights inherit their base model licenses; MiniMax-H3 LoRAs fall under the MiniMax H3 Community License.