RoboBrain 2.5
BAAI's embodied brain model (悟界·RoboBrain), Apache 2.0 at 4B and 8B, shipped in parallel NVIDIA and Moore Threads builds. It plans and points in 3D for a controller underneath.
Overview
BAAI is the Beijing Academy of Artificial Intelligence, a state-backed non-profit, and RoboBrain sits inside its FlagOpen family next to the FlagScale training framework and the FlagOS multi-chip stack. That parentage explains the model's shape. It is not a vision-language-action model that drives joints. It is the layer above: affordance prediction, spatial referring, trajectory forecasting, long-horizon planning across several robots and a running structured memory of the scene.
The 2.5 generation, described in arXiv 2601.14352, moved the spatial output from 2D pixel predictions to depth-aware 3D coordinates with absolute metric constraints, and added dense temporal value estimation so the model can score how far along a task is. That second capability is the one to watch, because it turns the brain into a reward model for training the policy underneath it. Checkpoints are on Hugging Face under BAAI: 8B-NV and 8B-MT from 9 January 2026, the 4B from 1 March 2026, with the older 2.0 family at 3B, 7B and 32B still up.
The China-specific detail worth knowing is the silicon. FlagOS has ported this family across a growing set of domestic accelerators, and Ascend builds of RoboBrain 2.0 are published under FlagRelease on Hugging Face while the 8B-MT variant was trained end to end on Moore Threads hardware. If your deployment has to run on chips you can legally buy in Shenzhen, that is not a footnote, it is the whole procurement plan. The weaknesses are documentation depth and pace. English docs are thin past the README, and with 2.0, 2.5, X0 and the newly announced Orca all in flight, anyone building on this should pin a revision hash and stay on it.
Key Features
- ✓ Apache 2.0 across the whole line, weights and code, with none of the acceptable-use riders or monthly-user caps common in this category.
- ✓ RoboBrain 2.5-8B released 9 January 2026 in two builds from identical data and architecture: 8B-NV trained on an NVIDIA cluster and 8B-MT trained on a Moore Threads cluster. The 4B followed on 1 March 2026.
- ✓ Built on a Qwen3-VL backbone, roughly 9B parameters in the published safetensors, BF16.
- ✓ Predicts depth-aware 3D coordinates with metric constraints rather than 2D pixel points, and adds dense temporal value estimation that can be used as a reward signal for VLA training.
- ✓ The 2.0 line stays available at 3B, 7B, 7B-FP8 and 32B (33B parameters), with FlagOS ports for Huawei Ascend published under the FlagRelease org on Hugging Face.
- ✓ RoboBrain-X0, a separate Apache 2.0 cross-embodiment VLA previewed 29 September 2025, covering AgileX, Franka, R1-Lite and A2D arms, with its training dataset published alongside.
Where it holds
- • The licence is boring in exactly the right way. Apache 2.0 on the weights means legal review takes an afternoon, not a quarter.
- • Ascend and Moore Threads builds actually exist and are published, which matters if your compute is domestic and export controls shape your procurement.
- • It is a brain, not a controller, so it sits above whatever low-level policy or motion planner you already trust instead of demanding you replace it.
- • Small enough to matter: the 4B runs on a single workstation card, so the planning layer does not have to live in a datacentre.
Where it breaks
- • The headline claim of 95 percent plus success from one demonstration comes from BAAI's own evaluation. No independent replication I can point to.
- • Release cadence is fast and confusing. RoboBrain 2.0, 2.5, X0 and the Orca model unveiled on 12 June 2026 all coexist, and Orca has no public checkpoint I can confirm. Pin a specific revision.
- • English documentation stops at the README. The FlagScale and Megatron training path assumes prior experience with both.
- • The 32B in BF16 needs roughly 70 GB just for weights, so the strongest variant is a two-card job before you touch context.
My Take
Depth-aware 3D coordinates instead of 2D points. That one change removes a downstream calibration step which used to eat a week, and it is the real upgrade over 2.0. The 2.5-8B release on 9 January 2026 shipped as two builds from the same data, one trained on NVIDIA and one on a Moore Threads cluster, which tells you plenty about where BAAI expects its users to be running. Everything is Apache 2.0, weights included, so unlike most embodied releases coming out of China you can put this in a product without a licensing negotiation.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0
- Starting at:
- Free. Weights and training code are on Hugging Face and GitHub under Apache 2.0 with no user cap, no field-of-use rider and no separate commercial licence to sign. Your only cost is the GPU time.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where RoboBrain 2.5 holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to RoboBrain 2.5
AgiBot World
open-sourceOpen manipulation data and policy weights from AgiBot (智元机器人): a million-plus real robot trajectories, the GO-1 VLA and the Genie Envisioner world model, all non-commercial.
Unitree UnifoLM
open-sourceUnitree's open robot learning stack: world models, a vision-language-action policy, Isaac and MuJoCo training environments and the SDK that drives Go2, G1 and R1 hardware.
SpatialLM
open-sourceManycore Tech (群核科技) of Hangzhou, behind Kujiale and Coohom, open-sourced its indoor spatial stack. SpatialLM turns point clouds into structured layouts, SpatialGen generates scenes.