Pinned Tweet
Join us in Shanghai for ModelScope DevCon 2026, an open-source AI gathering expected to draw around 5,000 attendees. Meet the founders, CTOs, and technical leads behind the models and tools developers build with. Tap the poster to see who's joining us.👇
📍Dec 12 | Hall N, West Bund Art Center, Shanghai
🔗 Register to join us at modelscope.ai/DevCon2026
Admission is free, and spots are limited.
CCORE-Web-Coder-4B turns a brief into a React HTML page you can open in a browser. This open-weight 4B preview is fine-tuned from Qwen3.5-4B-Base for local web generation. 📜 Apache 2.0.
🤖 modelscope.ai/models/ccore/C…
🖥️ Websites, slide decks, and informational pages. Generates a single HTML document with embedded CSS and JSX, using CDN-loaded React and Babel. No build step required.
📦 4.48 GB Q8_0 GGUF. Available for local inference, with examples covering portfolios, presentations, and reports.
🛠️ Use the supplied system prompt. Keep reasoning enabled and allow around 20K output tokens for the reasoning and complete HTML.
WeChat Vision at Tencent releases WeMM-Video, an open-weight model family for joint audio-visual understanding and reasoning. Available in 2B, 4B, 9B, and 35B-A3B sizes, it accepts text, images, audio, and video and responds in Chinese or English. 📜 Apache 2.0.
🤖 modelscope.cn/collections/We…
📄 Technical report modelscope.ai/papers/2611.26…
🎬 Understand the picture and the soundtrack together. Timestamped frames and audio tokens share a timeline, connecting actions, dialogue, and ambient sounds.
🏆 Strong audio-visual results. The paper reports SOTA among comparably sized open models for 35B-A3B. Its Thinking mode beats Seed2.0-Mini-0428 and Qwen3.5-Omni-Flash on 16 of 17 general audio-visual benchmarks each.
🪶 Competitive at smaller sizes. The 9B Thinking model outperforms Qwen3-Omni-30B-A3B-Instruct and MiniCPM-o-4.5-9B across most general audio-visual benchmarks in the paper.
🧠 Direct answers or deeper reasoning. Switch between Instruct and Thinking modes across the family.
🛠️ Choose your deployment format. BF16, FP8, AWQ, GGUF, and NVFP4 checkpoints, with Transformers support and patched vLLM/SGLang serving.
Hy4 is now open-weight. Tencent’s multimodal flagship has a 770B language backbone with 49B active per token, a 1M-token context, and text, image, and video understanding for coding, computer use, and research.
🤖 modelscope.ai/models/Tencent…
🏆 Strong agent results. Leads the evaluated open-weight models on OSWorld 2.1, Tencent’s internal TenureTrail research benchmark, and its 579-task finance evaluation.
🖥️ Vision meets execution. Combines visual understanding with CLI and GUI tool use, with targeted training for Blender, CAD, and workflows spanning multiple applications.
🎮 Built for game development. Adapted for Unreal Engine, Unity, and Godot, covering gameplay logic, state tracking, asset pipelines, and multiplayer development.
🛠️ Deploy and customize. FP8 weights, vLLM/SGLang serving, and official ms-swift scripts for full and LoRA fine-tuning on text and vision-language tasks.
RadixArk introduces DaisyRose-27B, an RL post-training release built on Qwen3.8-27B to make long-horizon agents more capable at real terminal work. Apache-2.0 License.
🤖 modelscope.ai/models/RadixAr…
🏆 More than doubles Terminal-Bench 4.1: Pass@1 rises from 5.6 to 12.0, the largest gain across its reported evaluation suite.
🛠️ Trained for long-running agent tasks: RL used 1,200 curated tasks spanning software engineering, science, and machine learning, with sessions reaching up to 1M tokens during training.
📈 Improves beyond terminal tasks: It also outperforms its Qwen3.8-27B base on Terminal-Bench 2.1, DeepSWE, SWE-Bench Pro, instruction following, and reasoning benchmarks.
🖼️ Keeps the base model’s multimodal stack: DaisyRose retains image understanding and a 256K-token context window, while reasoning is enabled by default.
⚡ Three deployment-ready checkpoints: Choose BF16, FP8, or NVFP4 weights, with SGLang and DSpark speculative decoding recommended for serving.
SenseNova-Loop-9B is out, an open-source 9B multimodal model that pushes deeper computation through shared layers. Apache-2.0 License.
🤖 modelscope.ai/models/senseno…
🏆 Big gains over Preview: AutomationBench rises from 14.33 to 36.67, Terminal-Bench 2.1 from 40.45 to 59.55, and GAIA from 37.86 to 69.90.
📊 9B, competitive with 27B: It matches Qwen3.8-27B on AutomationBench and reaches 65.31 on AIDABench, slightly ahead of Qwen3.8-27B’s 65.01.
🧠 Deeper computation, not more weights: Loop + mHC turns 32 distinct decoder layers into 56 layer executions per forward pass, reusing the middle-layer weights twice.
🖼️ Long-context multimodal: Accepts text and images, generates text, and supports up to 512K context for multi-step tasks, coding, and tool use.
💾 Built for smaller GPUs: Official 4-bit AWQ weights are about 4.9 GB; with a 4-bit KV cache, 64K-context deployment can fit in roughly 9 GiB of VRAM.
1
6
1,097
We are glad to see ms-swift included in supporting fine-tuning of Tencent Hy’s Hy4-flash on Day0. This official plugin from Tencent includes full and LoRA SFT recipes for both text-only and vision-language tasks.💻 github.com/modelscope/ms-swi…
As the open-source training framework from ModelScope, ms-swift has been facilitating customization of latest open models, and we welcome Hy4-flash as the latest addition in ms-swift family.
Tencent Hy has released Hy4-flash as an open-weight 220B MoE model that activates just 12B parameters per token, with native image understanding and 1M context.
🤖 modelscope.ai/models/Tencent…
🏆 Built for lower-cost agent work: On Tencent’s internal WorkIndex, Hy4 flash slightly outperforms the much larger Hy4 Preview while cutting per-task cost to 1/13 of Hy4 Preview and 1/20 of Hy4.
⚡ Made to serve at scale: It exceeds 150 tokens per second in decoding, while its KV cache is just 1/24 the size of Hy3’s and 1/6 of Hy4’s.
🧠 Long context without the usual overhead: Hybrid attention, LatentMoE, Engram, and the built-in DFly2 drafter support 1M-token workflows with lower, more predictable decoding cost.
🖼️ Three checkpoints, different jobs: Choose Hy4 flash or Hy4 flash-FP8 for instruction following and reasoning, or Hy4 flash-base for continued pretraining and customization.
🛠️ Ready for common runtimes: Production recipes are available for vLLM and SGLang, with Transformers support and an OpenAI-compatible API workflow.
📜 Tencent Hy License applies.
10
1,365
Tencent Hy has released Hy4-flash as an open-weight 220B MoE model that activates just 12B parameters per token, with native image understanding and 1M context.
🤖 modelscope.ai/models/Tencent…
🏆 Built for lower-cost agent work: On Tencent’s internal WorkIndex, Hy4 flash slightly outperforms the much larger Hy4 Preview while cutting per-task cost to 1/13 of Hy4 Preview and 1/20 of Hy4.
⚡ Made to serve at scale: It exceeds 150 tokens per second in decoding, while its KV cache is just 1/24 the size of Hy3’s and 1/6 of Hy4’s.
🧠 Long context without the usual overhead: Hybrid attention, LatentMoE, Engram, and the built-in DFly2 drafter support 1M-token workflows with lower, more predictable decoding cost.
🖼️ Three checkpoints, different jobs: Choose Hy4 flash or Hy4 flash-FP8 for instruction following and reasoning, or Hy4 flash-base for continued pretraining and customization.
🛠️ Ready for common runtimes: Production recipes are available for vLLM and SGLang, with Transformers support and an OpenAI-compatible API workflow.
📜 Tencent Hy License applies.
inclusionAI introduces Duo v1, its first implementation of a Reciprocal World Model, where the AI becomes a participant in a shared world instead of a system that only answers from outside it.
🤖 modelscope.ai/collections/in…
🌍 A world that responds back: Duo is designed for full-duplex, real-time audiovisual interaction, continuously perceiving what a person says and does while generating its own audio and video responses.
🗣️ Knows when to join in: Its Intent module reasons over the ongoing interaction and decides whether to speak or wait, enabling more natural interruptions, timing, and turn-taking.
🎥 Makes interaction visible: The Realization module turns those decisions into synchronized video and stereo audio, supporting shared activities such as showing objects, reshaping scenes, guidance, and audiovisual games.
📦 Four released checkpoints: Duo 1.0 includes causal-base, distilled, and quantized variants; Duo 1.2 adds a causal-base checkpoint with improved generation quality.
⚡ Built for research and streaming: An offline pipeline is available for research, while vLLM-Omni and SGLang-Omni support accelerated real-time deployment.
📜 Code is released under Apache-2.0. Model weights follow the MiniMax H3 Community License.
LongCat-Video-Avatar 1.6 is out, pushing audio-driven avatar video past 10 minutes with more natural speech motion and stronger long-form consistency. 📜 MIT License.
🤖 modelscope.ai/models/meituan…
🏆 Ranks #2 in human preference: It places second among six evaluated systems in a 33,051-vote Arena, ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5, and Kling Avatar 2.0.
🗣️ More natural talking avatars: The update improves lip-audio sync, facial and body motion, and reduces exaggerated mouth artifacts during speech.
🎬 Built for long-form video: It reduces error accumulation during autoregressive continuation to preserve identity, visual quality, and temporal coherence beyond 10 minutes.
🧩 More than a single portrait workflow: Supports single- and multi-character audio-driven video generation, including audio-text-to-video, audio-image-to-video, and video continuation.
⚡ Optimized for serving: Includes end-to-end inference improvements, task-specific LoRAs, and an optional weight-only INT8 DiT checkpoint.
Try a live, real-time voice conversation with Haru-4B on ModelScope. Demo 👉 modelscope.ai/studios/AI-Mod…
Upload or record a 3-second to 3-minute voice reference, set a character and conversation language, then start talking. Hold the live mic button to speak and follow the exchange with real-time captions. Haru is designed to keep the conversation flowing with natural backchannels and interruptions, rather than making you wait for each turn to end.
Headphones recommended. Each session supports up to five minutes of conversation.
What if your research agent didn’t stop at the first answer?
Built by @SolenopsisCampo , Venastine Research Harness is now available to try on ModelScope. Its ten-pass Research mode drafts an answer, checks claims against sources, challenges assumptions, and revises only what gets flagged.
Use Chat for everyday tool-assisted questions, or Research for work that needs a more auditable trail—all in a terminal-style workspace in your browser.
🤖 modelscope.ai/studios/Venast…
Edge0-145B brings 145B-scale agent capabilities to a single consumer GPU, with 10.12B active parameters per token and 256K context.
🤖 modelscope.ai/collections/Ed…
🏆Strong agent results: In Edge0’s reported comparison, the 145B model tops the listed models on Terminal-Bench 2.1 with 86.5, while reaching 64.6 on DeepSWE 1.1.
⚡ 145B, running locally: The Q3_K build reports 121.2 tok/s for agent decoding and 3,554.9 tok/s for 16K-token prefill on an RTX 5090 D 24GB.
🧠 Sparse by design: Its latent MoE architecture activates just 10.12B parameters per token across 512 experts, making a much larger model practical for local deployment.
🛠️ Pick the deployment fit: Edge0-145B prioritizes quality, while Edge0-120B offers a faster 8.64B-active alternative. Both are available in Q3_K and NVFP4 builds through Edge0’s dedicated local runtime.
📜 Edge0 Community Model License 1.0 applies to Edge0 contributions; upstream Qwen materials remain under the Qwen Community License 1.0.
RoboValue-Dataset is now on ModelScope: fine-grained training and evaluation data for robotic value models.
🤖 modelscope.ai/datasets/RoboV…
📦 Sim and real manipulation: 3,535 expert episodes across 15 simulation and 20 real-world tasks.
🧠 Evaluate execution, not just outcomes: Covers task state, progress, failure recovery, and value consistency.
🔍 Stress-test value judgments: Counterfactual instructions, repeated states, recovery trajectories, and alternative task orders expose unreliable feedback.
🔄 Test generalization: Includes cross-embodiment and cross-environment settings, with Full-Shot and One-Shot subsets in LeRobot v3.0 format.
📜 Code, scripts, and documentation are released under the MIT License.
Voder is now open source. Built by the VeriLoop team and supported by the Intelligent Robotics Laboratory at Tsinghua SIGS, it’s a coding agent that verifies proposed changes before writing them to the workspace.
🛡️ Independent checks before write-back. With VeriLoop-Governed Recurrence (VGR), the agent proposes and revises changes. A Rust governance layer authorizes the write-back only after the exact proposed change passes independent checks. The team is exploring how this approach could make recursive self-improvement safer.
📊 Locally audited benchmarks with DeepSeek V4.1 Flash 1017
• DeepSWE v1.1 · 78.8% (89/113)
• Terminal-Bench 3.0 · 41.9% (31/74)
• Terminal-Bench 4.0 · 45.5% (30/66)
• SWE-Marathon · 45.0% (9/20), with reference assistance
These are the team’s local audit results. Voder is available on GitHub and the VS Code Marketplace.
📄 modelscope.ai/posts/434415
💻 github.com/brucewang12345678…
🚀 MiniMax-H3-Image is now live on ModelScope Civision!
✍ Turn words into visuals
🖼 Mix & edit with up to 8 reference images
✨ Bring every detail to life in native 4K
One idea. Endless possibilities.
🔗 Create something amazing: modelscope.ai/civision/image…
Unsloth reports 5-10% gains across most tested benchmarks with Dynamic V3.1 (preview), its latest quantization update for the open-weight Qwen3.8-27B GGUFs. 📜 Apache-2.0 License.
🤖 modelscope.ai/models/unsloth…
🧪 Evaluated on practical workloads: coding, agents and chat, with benchmarks including Terminal Bench and DeepSWE.
🎛️ Pick your memory budget: most quantizations from 2-bit to 6-bit are available. The 3-bit files retain UD3 because V3.1 did not deliver enough improvement to justify replacing them.
💻 Run locally: the 4-bit version targets roughly 16-19GB VRAM, or a Mac with 24GB memory. Actual requirements depend on context length and offloading.
🛠️ Choose your stack: run through Unsloth Desktop, serve via unsloth run, or use llama.cpp.
📊 Benchmark details matter: Unsloth's analysis examines embedding-size accounting, context budgets and repeated runs, rather than relying on a single score.
2
25
3,037
The Sophon-Realtime demo is now available on ModelScope Studio. Try the latest version live! modelscope.ai/studios/Advent…
Sophon-Realtime v1.1 from Adventists AI @BKJinAI adds zero-shot voice cloning to an open-weights, full-duplex voice model.
Drop a few seconds of someone's voice into model/voices/, set "voice" to its name, and talk to it in that voice in real time. No fine-tuning, no restart, one GPU. Barge-in and laughs still work, and a voice cloned from a clip in one language can speak another.
Available in the Tag and Flash builds on ModelScope. Live mode is a drop-in replacement for the OpenAI GPT-Live API, and an E2E build with 0.21 s to first audio is coming soon.
Please only clone voices you have permission to use. 🤖modelscope.ai/models/Adventi…
1
16
2,095
Deploy MiniMax-H3-Image on ANY hardware with ultra-low cost! 🤖 modelscope.ai/models/DiffSyn…
Fast, low-cost inference is now possible with EntroPack. Choose your checkpoint:
🔹 80GB VRAM: bf16_table_pruned.safetensors (Fully lossless pruned model)
🔹 24GB VRAM: dit_8bit_int8.safetensors (Near-lossless INT8 + ConvRot quantization)
🔹 8GB VRAM: dit_extreme_3bpp.safetensors (Extreme 3-bit quantized model)
ComfyUI workflows also available in the model repository. Try it now!
Switch from Claude Code to Codex, Cursor or your own harness, and Nex-N2.6 still calls only the tools that are actually there. The open-weight 321B multimodal agentic model from Nex-AGI is built on GLM-5.3-Flash for computer use, coding and long-horizon tasks. 📜 Apache 2.0 🤖 modelscope.ai/models/nex-agi…
🛠️ Harness-agnostic tool use: 100% valid tool selection across 956 real-world contexts, matching GPT-6-Sol and up from 80.56% for Nex-N2.5-Pro.
🔁 Far less looping: reasoning repetition drops 90%, from 16.0% to 1.6%, on 698 real failure cases replayed from the exact point where the previous model stalled. Fewer wasted tokens on long runs.
🖥️ Stronger computer use: 83.0 on OSWorld, 5.0 points above its GLM-5.3-Flash base, while holding 84.1 on Terminal-Bench 2.1.
⚡ Ready to run: 256K context via API, or self-host the FP8 weights on SGLang behind an OpenAI-compatible endpoint with low, high or max reasoning effort.




















