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.
- Tweets 171
- Following 23
- Followers 2,447
- Likes 135
Can we trust what the AI benchmarks says? Can an agent system be smarter than its best model?
Join inclusionAI at #OpenSourceAIWeek 2026 for tech talks, happy-hour drinks, food, and Token Shots 🍸
Bay Area · RSVP: luma.com/xs9sn0lp
1
9
372
🥳 Meet our brand new model family, Ming-Image-0.1-Design!
1️⃣ Ming-Image-0.1-Design is a 6B text-to-image model for UI, infographics, posters, and other text-rich visual designs, while 2️⃣ Ming-Image-0.1-Design-Layer converts flattened graphics into 2–9 independently editable RGBA layers.
We would also like to introduce you 2 open-source Agent Skills: the Ling UI Design Skill and the Image-to-Editable-PPT Skill.
#inclusionAI #opensource #ImageModel
Download and try out now🤗:
🔗Hugging Face huggingface.co/inclusionAI/M…
huggingface.co/inclusionAI/M…
🔗ModelScope modelscope.cn/models/inclusi…
modelscope.cn/models/inclusi…
We’re open-sourcing the Ming-Image-0.1-Design family:
• Ming-Image-0.1-Design, 6B
• Ming-Image-0.1-Design-Layer, 6B
• Two open-source Agent Skills: the Ling UI Design Skill and the Image-to-Editable-PPT Skill
Ming-Image-0.1-Design ranks #1 among open-weight models on Artificial Analysis’s UI/UX Design leaderboard. 🧵
5
8
104
6,752
Make financial AI more accessible and verifiable➡️meet Ling-3.0-flash-Fin FinFIRST!✨
💡With open weights, teams can deploy Ling-3.0-flash-Fin privately, connect it to search, Python, databases and spreadsheets, and adapt it to their own financial workflows.
🔧FinFIRST is an open benchmark for financial search agents. V1 includes 123 expert-authored tasks, 701 atomic criteria and 12,300 rubric points for tracing evidence, sources and calculations.
#OpenSourceAI #inclusionAI #LLM #FinancialAI
🔗Ling-3.0-flash-Fin: huggingface.co/inclusionAI/L…
🔗FinFIRST dataset: huggingface.co/datasets/incl…
We’re open-sourcing Ling-3.0-flash-Fin, a finance-enhanced model for real-world workflows, and FinFIRST, an expert-built benchmark for financial search agents.
Two open releases, one goal: making financial AI more accessible and verifiable.
2
1
14
1,407
We are excited to introduce LLaDA-Image and LLaDA-Image-Turbo.
Highlights:
🎨 High-quality generation for photorealistic images, posters, ads, and bilingual typography
🪄 One 6B DiT unifies text-to-image generation and instruction-guided editing
Both backone and Image-Gen are diffusion models, trained in a unified framework.
⚡ Fast 2-4-step inference with LLaDA-Image-Turbo
4
20
150
9,159
High-quality text-to-image pre-training need not begin with image-text pairs.
👉LLaDA-Image builds its visual generative prior from images. Of ~220M cumulative generation-training samples, >90% use image-only supervision; paired data handles later language alignment.
Image-only data builds the visual world; image-text pairs connect language to it.
Both backone and Image-Gen are diffusion models, trained in a unified framework.
1
6
1,272
On Qwen-Image-Bench, LLaDA-Image scores 53.53 on the English track and 53.38 on the Chinese track, advance rankings on both among the open-source models listed in the technical report.
Generation and editing share one backbone. The same model family offers two deployment profiles: Base prioritizes full quality, while Turbo prioritizes inference speed.
#OpenSourceAI #dLLM #inclusionAI
Try out and explore LLaDA-Image, the open release includes Base and Turbo weights, training and inference code, and the complete training recipes:
🤗 Hugging Face: huggingface.co/collections/i…
💻 Code: github.com/inclusionAI/LLaDA…
📄 Technical Report: arxiv.org/pdf/2609.03796
1
10
961
🤔What if a few cheap models could pool their intelligence and beat a much stronger model?
💥Turns out they can.
Today we’re releasing PanelWise — models catching each other’s blind spots and combining their strengths. Check the details: github.com/inclusionAI/Panel…
2
15
2,213
InclusionAI retweeted
Robbyant R2 has been recognized by IDEA and K-Design Award. 🏅
For humanoid service robots, great design is not just how it looks — it is how safely, efficiently, and naturally it works around people.
Our R2 industry solution will be showcased at WRC this week. 👋 Come meet us at C111.
2
3
12
1,078
We adapted the public DSpark recipe to @AntLingAGI with distribution aligned data, architecture ablations, and an acceptance aware loss.
Meet 🆕Ling-3.0-flash-dspark, a DSpark draft model built specifically for Ling-3.0-flash.
🔗Full engineering details: lmsys.org/blog/2026-08-21-li…
🔗Hugging Face: huggingface.co/inclusionAI/L…
🔗ModelScope: modelscope.cn/models/inclusi…
Today we are open sourcing Ling-3.0-flash-dspark, a DSpark draft model built specifically for Ling-3.0-flash.
On 4 NVIDIA Blackwell GPUs at batch 1, it delivered 1,120 tok/s, 0.78 ms mean TPOT, and an accept length of 9.95 across 1,000 requests. 🧵
1
1
10
1,141
InclusionAI retweeted
What made cold starts so slow? 🔍
Weight loading dominated the startup path. Each TP rank had to read around 120 GB from disk, deserialize it, apply TP sharding, run FP8 quantization, and repack the weights.
Every restart repeated the same deterministic work.
5
1
13
2,500
InclusionAI retweeted
🧵 We’ve open-sourced 6 Base Model checkpoints for Ling-3.0-tiny & Ling-3.0-flash, covering pre-trained, mid-trained, and WSM-merged stages. None has undergone post-training, giving researchers flexible starting points for continued pre-training, fine-tuning, and further research.
Two key highlights:
- We use WSM to replace LR decay with weighted checkpoint merging, making the training process better suited for continual pre-training while enabling offline exploration of different LR decay strategies.
- With one shared training recipe, the community can validate strategies on tiny-base, then scale them to flash-base.
22
43
360
521,685
AI proves its value through real tasks and real users. We want more people to be able to build with AI and afford to use it. Joining the PyTorch Foundation is another step in Ant Group's long-term participation in global open-source collaboration. We look forward to working with the community to lower the barriers to AI development and adoption through open models, infrastructure and agent technologies.🤝
🎉Ant Group has joined the PyTorch Foundation as a Gold Member. Through @TheInclusionAI, Ant Group works on open models, infrastructure and agent technologies that make AI easier to build and use. It also helped launch and continues to invest in AReaL @AReaL_AI, a PyTorch Ecosystem Landscape project. Welcome, @ant_oss! #PyTorch #OpenSourceAI #inclusiveAGI
5
1
21
1,068
🤔Can AI agents improve quant research without gaming backtests?
Meet AQuA: autonomous research loops🔁 in sealed sandboxes, learning from validated evidence.
US equities: 0.0843 IC, up to 2.50 held-out Sharpe @ 2bp, positive each year, 2021-2025.
👀Check the details: arxiv.org/abs/2608.12841
5
157,563
Lightweight models are moving from cloud APIs into local devices and real workflows.
🤔But after a model can run locally, can it also train locally?
Introducing AReno @ARenoTeam, which we used to post-train @AntLingAGI Ling-3.0-tiny on DGX Spark with an Agentic RL tic-tac-toe task.
The model interacts with an environment, calls tools, receives rule-based rewards, and updates its behavior with GSPO.
#inclusionAI #OpenSource #ReinforcementLearning
1
8
700
Results after 400 steps:
✅rollout/rewards_mean: ~-0.5 → ~0.4
✅response_len: down to ~850 tokens
✅fewer invalid actions
✅more stable tool calls and decisions
1
2
248
Tic-tac-toe is just the minimal verifiable task.
🔄The same loop can extend to tool-call repair, structured extraction, workflow agents, domain instruction following, and more.
From running, to training, to adapting to your own task.
GitHub: github.com/inclusionAI/AReno
1
135