ModelBest and OpenBMB Open-Source the MiniCPM-Robot Embodied AI Model Series

Core Highlights

ModelBest, together with OpenBMB, launched MiniCPM-Robot, the world's first open-source embodied AI model series, releasing the vision-language-action capability that previously lived only in papers directly as commercially usable open-source weights. This means robot developers no longer have to train from scratch to obtain a base model that can see, understand, and act, and it also means the research barrier to embodied intelligence is, for the first time, truly open to small and mid-sized teams rather than locked inside a few well-funded labs that can afford the compute to pretrain such models. By open-sourcing the weights, the release turns a research breakthrough into a shared community asset that anyone can download, inspect, and improve, which is a meaningful shift in how embodied intelligence is likely to be built going forward. The symbolic weight of the word 'first' here matters: it signals that the open community, not just the closed frontier labs, can set the pace in this emerging field.

What the Capability Actually Does

The series contains two core models. The first is MiniCPM-RobotManip, a general-purpose vision-language-action (VLA) model with 1.5B parameters, focused on manipulation and operation tasks for robotic arms, able to understand natural language instructions and map them to concrete grasping and movement actions. The second is MiniCPM-RobotTrack, dedicated to object tracking, letting a robot continuously lock onto and follow a specific object as it moves through space. Released alongside the models is a high-performance inference framework named PhyAI, whose goal is to provide robots with an integrated runtime foundation spanning perception, understanding, memory, and action, folding scattered capabilities into one deployable pipeline rather than a pile of disconnected scripts that a developer must wire together by hand. Together, the two models plus the framework form a starter kit: a team can take the weights, run them through PhyAI, and have a robot that perceives, reasons, and moves without having to assemble the stack from unrelated research repositories.

Technical Details

The key to a VLA model is unifying visual input, language understanding, and action output into the same model parameter space, so that 'what is seen, what is heard, and what should be done' are connected end to end instead of handled by three separate modules that must be carefully synchronized. The deliberately chosen 1.5B scale keeps the model within the range that can run on the edge, and together with the PhyAI inference framework it compresses large-model capability into the compute constraints of the robot body itself, without depending on a round trip to the cloud. Object tracking, in turn, gives the robot a persistent spatial attention, an indispensable prerequisite for any complex manipulation task that must keep its target in view while the arm moves, because a tracker that loses the object mid-action makes the whole manipulation fail. The design choice to split manipulation and tracking into two models also keeps each one small and fast, which suits the tight latency budgets of physical robots.

Comparison with Competing Approaches

Compared with approaches such as Google's RT series and Stanford's VIMA that are mostly papers and closed weights, MiniCPM-Robot's open-source strategy sharply lowers the barrier to entry for embodied intelligence. It continues the MiniCPM family's 'small parameters, run on the edge' line, giving it a stronger deployment advantage on carriers like robots that are sensitive to size, power, and real-time performance, and it also lets domestic researchers innovate directly on top of an open base instead of rebuilding from zero every time they want to test a new idea in the lab. Where closed models force you to trust a vendor's API and accept its constraints, the open weights let a team audit behavior, patch failures, and tune for their specific hardware, which is often the deciding factor in safety-critical robotics work.

Industry Impact and Use Cases

Simply put, open-source embodied models pull robot development back from 'competing on papers and compute' to 'competing on applications.' Educational research, lightweight service robots, and desktop robotic-arm developers can all build directly on top of it for secondary development, accelerating the speed at which embodied intelligence moves from the lab into real scenarios, and giving the country an early position in the open-source embodied-intelligence ecosystem, around which toolchains and communities will subsequently grow and reinforce one another. The practical upshot is that a university lab or a hardware startup can now prototype a capable robot without a multimillion-dollar training budget, which broadens the field beyond the handful of giants that previously dominated it. The open release also creates a feedback loop that closed models cannot match: as more teams deploy MiniCPM-Robot and report what breaks in the real world, the community can collectively patch the weights, and every downstream user benefits from fixes they did not have to write themselves. For a domain as physically grounded and safety-sensitive as robotics, that collective debugging is not a nice-to-have but a prerequisite for trust, because a robot that fails in the lab is an inconvenience while a robot that fails in a home is a hazard. In the longer arc, the move suggests that the center of gravity for embodied intelligence may shift from a few closed labs toward a broad, permissively licensed commons, where progress is measured less by a single flagship demo and more by the steady accumulation of small, reusable improvements contributed by many hands working on many different machines.