InclusionAI

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/Zo…
🤗huggingface.co/collections/in…

⚡ 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/Li…
🔗 Ling-3.0-flash-124B: huggingface.co/inclusionAI/Li…
Use them to accelerate your Ling-3.0 now!
More draft models are coming over the next month.
📄 Paper: inclusion-ai.org/ubtree/UBTree.…
🏠 Website: inclusion-ai.org/ubtree

⚡️⚡️⚡️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/Li…
ModelScope: modelscope.cn/models/inclusi…
🔗 Ling-3.1-flash-base
HF: huggingface.co/inclusionAI/Li…
ModelScope: modelscope.cn/models/inclusi…

🤗 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/Li…
🔗 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👀