←
InclusionAI191 posts
InclusionAI avatar

InclusionAI

@TheInclusionAI

AI Lab @AntGroup, we envision AGI as humanity's shared milestone. Our Language Model @AntLingAGI and LLaDA, Embodied AI @robbyant_brain, OSS projects @AReaL_AI.

◷ Joined March 2025▧ 35 Photos and videos
InclusionAI avatar
InclusionAI@TheInclusionAI·5h

InclusionMed launches as the first independent evaluation and analysis platform for medical AI, co-developed by 29 institutions including Stanford, Harvard, Peking Union Medical College and Ant Group. In InclusionMed, we evaluates not just how models score, but who they serve and what they can actually do. The platform introduces two tools: · Atlas: a nine-cell benchmark map that organizes evaluations by audience and capability. · Index: a composite score aggregating results from its first evaluation cycle of 1,034 tasks and 22 models. Homepage:inclusionmed.ai/ GitHub:github.com/panuvi/inclusionm…

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·23h

Duo v1 performance on 30 real-time duplex interactions (180 checks): • Language, speech, identity: 100% • A/V emotional consistency: 96.7% • Content responsiveness: 73.3% • End-to-end latency: ~708ms Try out and let us know what you build! 🔗arXiv: arxiv.org/abs/2611.15131 🔗Hugging Face: huggingface.co/collections/i… 🔗ModelScope:modelscope.cn/collections/in… 🔗 Homepage: duo-v1.github.io/ 🔗 GitHub: github.com/inclusionAI/Duo

InclusionAI avatar
InclusionAI@TheInclusionAI·23h

In RWM, we identify two equally fundamental capabilities: 1️⃣ Intent, which provides the cognition underlying world interactions. 2️⃣ Realization, which generates coherent audiovisual behavior. 👉 Duo v1 introduces three key designs: ✅ Autoregressive modeling for continuous interaction reasoning and participation decisions—not restricted to instruction understanding. ✅ Chunk-specific conditions with synchronized tokens across Intent and Realization on a unified timeline. Causal masks enforce dependencies within and across chunks, blocking future access. ✅ Dual rollout distillation mitigates distribution shift by training on autoregressively generated Intent tokens and synthesized audiovisual histories. Both modules perceive real-time audio, video, and text. Full-duplex interaction. Spontaneous conversations. Interruptible.

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·23h

🤔Most world models simulate physics and leave intelligence on the outside. We think differently. Intelligence and physical reality are equally fundamental to a world. We built something different, meet Duo, the first Reciprocal World Model, treating intelligence and physical reality as equally fundamental: a virtual presence that perceives, decides, and responds in real time natively. #WorldModel #OpenSource #ReciprocalWorldModel

1:04
InclusionAI avatar
InclusionAI@TheInclusionAI·Nov 10

Introducing LingBot Touch 🔥 A JEPA based tactile foundation model for robotic manipulation is now open sourced! 🏢→ A shared, scalable architecture supporting cross-sensor, cross-hand, cross-task, and cross-action-model transfer 🙌→ Real-robot validation across two dexterous hands, 26 manipulation tasks, and four action models 📊→ Tiny and Small variants for lightweight, high-frequency tactile inference 💥→ Nearly 5,000 hours of tactile pretraining data, with performance improving as the data scales.

What changes when robots can feel—not just see? 🤖✋ Meet LingBot-Touch, our open-source tactile foundation model towards dexterous manipulation, pretrained on nearly 5,000 hours of real tactile data. Through unified heterogeneous tactile modeling, LingBot-Touch supports transfer across sensors, dexterous hands, tasks, and action models. Watch it in action. 👇 #LingBot #EmbodiedAI #Robotics

1:37
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 29

Seeing details is only the beginning. Connecting them is the next step. 🚀 Introducing Zooming without Zooming v2 (ZwZ2). Our EvidenceConnect framework generates questions that require connecting visual evidence across regions within a single image. Trained with RLVR on just 20.4K of these examples, ZwZ2-4B/9B/27B improve over their corresponding base models on all 25 benchmarks: • ZwZ2-4B: 49.41% → 54.79% • ZwZ2-9B: 51.63% → 56.59% • ZwZ2-27B: 55.52% → 60.51% The gains extend to multi-image long-context understanding, embodied spatial reasoning, and multimodal agent tasks—even with single-image training. 📄 arxiv.org/abs/2610.37037 🛠️ github.com/inclusionAI/Zoomi… 🤗huggingface.co/collections/i…

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 29

⚡ Introducing UBTree: a frontier dLLM-based method for inference speedup🚀 On Ling3-Flash-124B, UBTree delivers ~1,010 output tokens/s with 4-GPU tensor parallelism (TP4) at concurrency 1, averaging a 3.43× end-to-end speedup over target-only autoregressive decoding across six benchmarks. UBTree uses a unigram proposer to provide candidate draft tokens and a bigram selector to connect these candidates into a tree. We build this draft tree completely in parallel. Today we are releasing two drafters for the @AntLingAGI family models: 🔗 Ling-3.0-tiny-8B: huggingface.co/inclusionAI/L… 🔗 Ling-3.0-flash-124B: huggingface.co/inclusionAI/L… Use them to accelerate your Ling-3.0 now! More draft models are coming over the next month. 📄 Paper: inclusion-ai.org/ubtree/UBTr… 🏠 Website: inclusion-ai.org/ubtree

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 27

⚡️⚡️⚡️Ling-3.1-flash-1019 is now open source! Further post-training for reasoning and real-world agent workflows. Along with another release, Ling-3.1-flash-base, an open foundation for further training. Download, evaluate and build with @AntLingAGI. 🔗 Ling-3.1-flash-1019 HF: huggingface.co/inclusionAI/L… ModelScope: modelscope.cn/models/inclusi… 🔗 Ling-3.1-flash-base HF: huggingface.co/inclusionAI/L… ModelScope: modelscope.cn/models/inclusi…

Introducing @AntLingAGI Ling-3.1-flash-1019 🚀 It scores 44 on the Artificial Analysis Intelligence Index, up from 41 for Ling-3.1-flash. Gains in agentic work: AA-Briefcase ~1400→1524 Elo and AutomationBench-AA 62%→68%. API access and open weights coming soon. Stay tuned👀

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 22

🤗 Meet our healthcare-enhanced MoE model, @AntLingAGI Ling-3.0-flash-Sante! Built for real-world healthcare research and workflows, with strengthened capabilities in deep research, multi-agent collaboration, and tool use. Sante shows a significant improvement on professional healthcare tasks, and provides stronger complex medical reasoning. 📊 Open source and open weights: 🔗 huggingface.co/inclusionAI/L… 🔗 modelscope.cn/models/inclusi…

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 22

Introducing @AntLingAGI Ling-3.1-flash-1019 🚀 It scores 44 on the Artificial Analysis Intelligence Index, up from 41 for Ling-3.1-flash. Gains in agentic work: AA-Briefcase ~1400→1524 Elo and AutomationBench-AA 62%→68%. API access and open weights coming soon. Stay tuned👀

Post attachment
SGLang avatar
⟳ InclusionAI retweeted
SGLang@sgl_project·Oct 19

SGLang-Omni supports @TheInclusionAI's Ming-3.0 on day 0! 🎙️ Ming-3.0-asr-tiny — Chinese/English ASR with dialect support 🔊 Ming-3.0-tts-tiny — voice cloning in 7 languages 🧠 Ming-3.0-audio-flash — 126B-A6.5B MoE for audio understanding and reasoning Cookbooks👇

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 20

Meet the Ming-3.0 speech families: ✅ Ming-3.0-tts for stories, narration & voice assistants ✅ Ming-3.0-audio for meetings, interviews & voice-driven services, offering a more efficient deployment option ✅ Ming-3.0-asr for customer calls, voice search & domain terms. 4.41% average error across four ASR categories. Try out on Hugging Face & ModelScope! huggingface.co/inclusionAI #opensource #inclusionAI

Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families. - Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech. - Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue. - Ming-3.0-asr: Tiny 8.6B A2B dialects & context.

Post attachmentPost attachmentPost attachmentPost attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 15

Build for real work, Ling-3.1-flash is now open source. ~560B total params, ~25B active/token, up to 1M-token context. Across work, ⌨️coding & 💪healthcare: 1,673 Elo on GDPVal-AA v2.1, 75.16 on FrontierSWE, and 65.35 on HealthBench Professional. 🤗Hugging Face: huggingface.co/inclusionAI/L… 🤖ModelScope: modelscope.cn/models/inclusi… #opensource #inclusionAI #LLM

Ling-3.1-flash is now open source: ~560B total parameters, ~25B active per token and up to 1M context. Built for real work, it can turn a rough brief into a page, mobile UI or interactive prototype teams can review and refine. See the demo.

1:43
Robbyant avatar
⟳ InclusionAI retweeted
Robbyant@robbyant_brain·Oct 7

The FIRST Robbyant Embodied AI Challenge is live. 🤖🌍 We're inviting builders worldwide to push embodied intelligence forward — real hardware, real environments, real impact. 🦾 Be part of history in the making. Prizes. Recognition. The frontier. 🏆 Learn more & register → tianchi.aliyun.com/competiti… #EmbodiedAI #Robotics

Post attachment
InclusionAI avatar
InclusionAI@TheInclusionAI·Oct 10

Speaker lineup’s here 👀 More names are on the way! Join InclusionAI Tech Night during #OpenSourceAIWeek for candid talks on open models, AI benchmarks, and real-world agents, plus Token Shots and networking! 📍 San Francisco 🗓 Oct 18 · 4–8 PM 🎟 luma.com/xs9sn0lp

Post attachment