prithivMLmods/Alibaba-DAMO-Academy_RynnBrain1.1-2B-GGUF

RynnBrain1.1-2B-GGUF

[RynnBrain1.1-2B](https://huggingface.co/Alibaba-DAMO-Academy/RynnBrain1.1-2B) is the smallest checkpoint in Alibaba DAMO Academy's RynnBrain 1.1 family of open embodied foundation models, built on Qwen3.5-2B and released alongside 9B and 122B-A10B (the family's first sparse-MoE model) variants under a unified training recipe designed to systematically study how embodied cognition, spatial reasoning, grounding, and planning evolve with scale. As part of this generation's upgrades, it introduces native 3D and contact-point grounding — extending beyond image-plane localization to metric 3D understanding via explicit 3D-grounded training and a new instruction-conditioned contact-point prediction task — and bridges perception to action through RynnBrain-VLA, translating embodied understanding into real-robot control with demonstrated cross-platform generalization on the Unitree G1, Astribot, and Tianji-Wuji platforms across humanoid, bimanual, and dexterous-hand tasks. The model handles spatial and object understanding, object/area/affordance/trajectory grounding, and both image and video inputs, outputting structured coordinate predictions (e.g., <object>(x1,y1),(x2,y2)</object> bounding boxes) for tasks like locating objects in a scene, and supports inference via Hugging Face transformers or SGLang (OpenAI-compatible serving or offline engine), with training and evaluation details documented separately in the companion RynnScale repository; it is released under Apache-2.0.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp