OpenBMB Open-Sources MiniCPM-Robot: 1.5B VLA Beats π0.5 by 5x on Memory Benchmarks
Source: OpenBMB (ModelBest) / X / media coverage URL: https://x.com/OpenBMB/status/2078839529591759025 Date: July 19, 2026 (at WAIC 2026)
The Release
On July 19 at WAIC 2026, Tsinghua-affiliated unicorn OpenBMB open-sourced its first embodied AI model series, MiniCPM-Robot, as a three-piece bundle:
- MiniCPM-RobotManip: a 1.5B-parameter general vision-language-action (VLA) model for robot manipulation.
- MiniCPM-RobotTrack: a 0.9B-parameter end-to-end visual instruction-following model for target tracking.
- PhyAI: a high-performance inference framework designed for embodied models.
- Real-robot generalization at 1.5B — benchmark numbers are strong, but RMBench is a new benchmark; performance under complex lighting, occlusion, and deformable objects remains to be seen.
- Open source ≠ ecosystem — Llama took two years to build an ecosystem; VLA models involve hardware adaptation, data collection, and sim-to-real, making ecosystem building far harder than for language models.
- Data provenance undisclosed — a 1.5B model typically requires millions of trajectories; their origin and compliance aren't clearly stated in the repo.
- No in-house robot manufacturing — OpenBMB lacks mass-production capability; whether the open-source model becomes an ecosystem play or relies on downstream integrators is unclear.
- Announcement: https://x.com/OpenBMB/status/2078839529591759025
- GitHub: https://github.com/OpenBMB/MiniCPM-Robot
- Hugging Face: https://huggingface.co/openbmb/MiniCPM-RobotManip
- Additional coverage: https://www.163.com/dy/article/L27AB1VS0511ABV6.html , https://www.chinaz.com/ainews/29702.shtml
All weights and code are on GitHub at OpenBMB/MiniCPM-Robot, with model weights also on Hugging Face. No review gates, no API quotas, no NDA — a rare level of openness in the 2026 embodied AI field.
The Key Isn't 1.5B — It's Memory
The pain point of embodied AI isn't perception; it's contextual memory.
Most mainstream VLAs (π0.5, RT-2, OpenVLA) are "single-frame reactive": they see the current frame, decide the next action, then discard everything. Ask one to press a button 5 times, and by the second press the model has already forgotten the first.
Why not just add history? Compute can't keep up. Each additional historical frame causes near-explosive growth in computation, and forcing memory in makes the robot sluggish — too slow to be usable.
MiniCPM-RobotManip's solution is characteristic of ModelBest: porting MiniCPM-V 4.6's visual token compression techniques to the robotics setting. The results:
| Metric | π0.5 (mainstream) | MiniCPM-RobotManip | Ratio | |---|---|---|---| | RMBench context-memory score | 10.4 | 53.5 | 5.1x | | Compute per decision step (60-frame history) | 5.0 TFLOPs (single frame) | 3.3 TFLOPs (streaming) | -34% | | H100 single-decision inference latency | 234ms | 120ms | -49% |
Behind these numbers is streaming inference: no recomputation of historical frames, only incremental updates to a visual token cache. As a result, "the inference cost of keeping 1 minute of context memory" is on par with traditional single-frame reactive approaches. Better performance and lower deployment cost hold simultaneously — a rare combination in the 2026 VLA landscape.
RobotTrack: 0.9B on a Single Card, Tracking Even in Offline Elevators
MiniCPM-RobotTrack set open-source SOTA across three tasks:
| Task | MiniCPM-RobotTrack (0.9B) | vs OmTrackVLA | vs TrackVLA++ (closed) | |---|---|---|---| | Single-target tracking (STT) | 89.8 | +7.0pp | +8.8pp | | Dynamic multi-target (DT) | 73.4 | +14.6pp | — | | Ambiguous-target (AT) | 80.4 (58.0% success rate) | +6.5pp | +17.0pp |
It natively supports the Unitree Go2 Edu, running at 5+ Hz with 180ms end-to-end response on stock cameras and onboard compute. It has been demonstrated smoothly tracking in "offline/weak-network" scenarios like elevators and parking garages — arguably more significant than the benchmark numbers: real-robot deployment pushed down to consumer-grade hardware.
Industry Context: 1.5B Isn't "Small," It's "Real-Robot-Runnable"
| Model | Params | Real-robot requirement | |---|---|---| | π0.5 (Physical Intelligence) | ~3B | H100/A100 | | OpenVLA-7B | 7.8B | Multi-GPU | | RT-2-X | 55B | Massive cluster | | MiniCPM-RobotManip | 1.5B | Consumer GPU |
ModelBest has pushed "embodied AI democratization" from paper demos to real robots that individual developers can clone and run. Once the barrier is kicked open by open source, the pace of iteration in embodied AI may exceed most expectations.
Why It Matters
1. Memory is the bottleneck of embodiment — while most teams try to route around it, ModelBest attacks it head-on with visual token compression. Genuine hard tech, not a wrapper. 2. Full 1.5B weights released — rarely seen openness in embodied AI, letting researchers, hobbyists, and startups outside big labs get hands-on. 3. Deployment threshold pushed to consumer grade — a ~10k RMB Unitree Go2 dog plus one consumer GPU can run end-to-end VLA. The "iPhone moment" of embodied AI may be closer than expected. 4. Same-day WAIC momentum — Riemann Dynamics' Riemann-1.0 (232k hours of embodied interaction data) and Xu Huazhe's panel on an embodied "ChatGPT moment" signal the embodied track collectively taking off in 2026 H2.
Risks and Open Questions
Bottom Line
1.5B beating π0.5 by 5x, 120ms on a single card, 180ms tracking on a Unitree Go2 — ModelBest has dragged embodied AI from "lab demo" to the tipping point where developers can clone and run. That makes it more worth tracking than LingBot-VA's more complex VAE architecture, because it bets directly on "real-robot runnability."