ModelBest and OpenBMB Release Open-Source MiniCPM-Robot Series with 1.5B VLA and 0.9B Tracking Models
Core Highlights
ModelBest, together with OpenBMB, released and open-sourced the MiniCPM-Robot series, including the general VLA (vision-language-action) model MiniCPM-RobotManip (1.5B parameters) and the mobile tracking model MiniCPM-RobotTrack (0.9B parameters). This is an important fill-in for end-side robotics large models, compressing embodied intelligence that once required a data center into a lightweight scale runnable on local devices, and doing so under a license that lets anyone reproduce and extend the work without asking permission from a vendor.
The release arrives at a moment when robotics is constrained less by ambition than by the cost and power budget of the "brain" that drives a machine. By fitting capable models into a footprint that runs on commodity edge hardware, the team lowers the economic threshold for putting intelligence into physically small and cheap products, from classroom arms to delivery bots.
What Happened
MiniCPM-RobotManip is positioned as "general manipulation," responsible for letting a robotic arm understand language commands and visual scenes, then plan and execute actions such as grasping and placing; MiniCPM-RobotTrack focuses on "mobile tracking," keeping the robot locked onto a target object or person for following and obstacle avoidance. With only 1.5B and 0.9B parameters respectively, they accomplish embodied tasks that vision-language models of tens of billions of parameters once hesitated to touch, suggesting ModelBest made targeted trade-offs in model compression and robot-specific architecture rather than simply shrinking a general model.
The whole series is open-sourced, meaning developers can fine-tune and deploy it on their own hardware without vendor lock-in or recurring cloud fees that make small robots uneconomic. That openness is what turns a research result into a platform: a university lab or a startup can take the weights, adapt them to a specific gripper or chassis, and ship a product without negotiating licenses for every deployment.
Technical Details
The difficulty of a VLA model is unifying "see, say, do" into one reasoning chain: the model first encodes image and language, then outputs action tokens or control signals. For 1.5B parameters to be usable, it needs careful design in data recipe (simulation plus real trajectories), action representation (discrete vs continuous regression) and distillation strategy, to avoid a small model forgetting its goal in long-horizon tasks where attention drifts.
The 0.9B RobotTrack is more like a real-time perception module, maintaining stable tracking of the target at low latency, fitting the compute constraints of embedded cameras and mobile chassis where every millisecond and megabyte counts and a heavy model would simply miss the moving target. Treating tracking separately from manipulation also lets each model be sized to its actual job rather than both inheriting the overhead of one giant network.
Comparison with Competitors
Compared with robotics large models such as RT-2 and OpenVLA that run into tens of billions of parameters, the salient tag of MiniCPM-Robot is "small and open-source." Larger models have higher capability ceilings but are hard to run in real time on the edge; ModelBest chose a smaller footprint for deployability, fitting the domestic robot need for "low cost and fast landing."
The open-source strategy also lets universities and startups access it with zero barrier, accelerating the democratization of embodied intelligence instead of reserving it for well-funded labs with clusters to spare and the engineering headcount to operate them in production. When the best robot brains are free to download, the bottleneck shifts from access to integration—a problem the broader community is well equipped to solve.
Industry Impact and Use Cases
Simply put, whether robots can spread widely hinges on whether their "brain" is cheap and can run locally. MiniCPM-Robot compresses VLA and tracking to the edge, serving education robots, warehouse sorting, home care and agricultural inspection while reducing dependence on the cloud and high-end GPUs. For the domestic hardware ecosystem, such lightweight open-source models are the glue connecting "domestic chips plus domestic robots," promising to lower overall solution cost and push embodied intelligence from the lab onto the production line.
As more roboticists build on the release, a local toolchain for affordable embodied AI becomes far more plausible, and the cost curve of smart machines can finally bend downward for smaller buyers. The strategic payoff is a robotics stack that does not depend on foreign cloud services—an increasingly important property as supply chains and export rules grow more uncertain.
The bigger picture is a domestic robotics stack taking shape from chips to models. When the brain, the body and the silicon are all sourced locally and tuned together, integration gets easier and supply risk drops, giving Chinese robot makers a clearer path from prototype to mass production than peers dependent on foreign components and overseas cloud services. Self-reliance in robotics is thus as much an engineering choice as a strategic one, with consequences that extend well beyond any single product launch. The cumulative effect is a more resilient domestic supply chain for intelligent machines.