MiniCPM5-2B Tops Sub-4B Models Worldwide and Adapts to Nine Chips

Core Highlights

ModelBest, together with OpenBMB, released the on-device model MiniCPM5-2B, which with only 2B model complexity took the top score of 17 points and an average of 54.26 on the sub-4B group of the AA-Index leaderboard, directly surpassing a host of competitors including Qwen3.5-2B. In other words, it delivers stronger on-device performance than larger models while using far fewer parameters, and it confirms the claim that a small model can still top a leaderboard rather than being dismissed as a toy that is only good enough for trivial tasks on a phone. The result matters because the sub-4B category is exactly where most real edge hardware lives, so leading there means leading where the volume shipments actually are, not just in benchmark bragging rights on a server GPU.

What the Capability Actually Does

One of the core selling points of MiniCPM5-2B is native support for hybrid thinking, meaning the model can switch on demand between 'fast thinking' and 'slow thinking': it answers simple tasks directly and only expands into multi-step reasoning for complex tasks, balancing latency against quality. At the same time, it natively supports a 512K context length, meaning it can swallow an entire long document or a whole code repository at once without breaking a sweat. Even more importantly, the model has completed Day0 adaptation on nine mainstream chips including Huawei Ascend and NVIDIA, so it runs out of the box across different hardware without waiting for the community to port it slowly, which is what usually blocks edge deployment of new models and keeps useful research stuck on a single vendor's board. The combination of a small footprint, a long context, and broad chip support is what makes this model feel less like a research release and more like a product-ready component a hardware team can ship.

Technical Details

Day0 adaptation means usable support is provided on the very day a new chip launches, rather than waiting for the community to port it slowly over the following weeks and months. The 512K context relies on long-sequence attention and memory optimization, letting on-device hardware also handle extremely long inputs that would normally overflow memory on a small device. Hybrid thinking is typically implemented by injecting thinking tokens during training, giving a single set of weights two answering styles at once, and at inference time the model picks the style based on how hard the task is, so easy requests stay snappy while hard ones still get the depth they need. These are not three independent tricks but one coherent design philosophy: spend parameters only where they earn their keep, and spend compute only when the task demands it.

Comparison with Competing Approaches

Compared with same-tier models such as Qwen3.5-2B, the advantage of MiniCPM5-2B lies in the combination of 'small enough to run on device, strong enough to top leaderboards.' Rather than chasing parameter scale, it pushes cost-performance and hardware compatibility to the extreme, which is especially friendly to resource-constrained scenarios like phones, car infotainment systems, and development boards, and it means the same experience can be rolled out cheaply to a massive number of devices instead of only to well-equipped servers that few users can afford to run. In a market where most AI still runs in someone else's data center, a model that is both small and broadly portable is a genuinely different bet on where the value will accumulate.

Industry Impact and Use Cases

Simply put, this kind of high-efficiency on-device model brings large models out of the server room and into the pocket. It will soon be open-sourced, so developers can embed it into phones, edge gateways, and smart cockpits for offline inference, protecting privacy while saving cloud-call costs, and it is a key link in the real scaling of on-device AI, as well as a ready-made template for the software-hardware co-design of domestic chips that want a proven model to run on day one. The strategic point is that every device that ships with a capable local model is a device that no longer needs a constant, paid connection to a foreign cloud, which is both a cost story and a sovereignty story for the local AI stack. When inference happens on the device, the raw inputs, whether they are a private document, a family photo, or a customer record, never leave the hardware, so the privacy guarantee is structural rather than promised in a terms-of-service paragraph. That property is what makes on-device models attractive not only to consumers who dislike surveillance, but to regulated industries that are simply not permitted to send certain data off-premise. And because the model is open, a chip vendor can co-design the next silicon generation around its exact operators, tightening the loop between algorithm and accelerator until the whole stack behaves as one tuned system instead of a model awkwardly ported onto whoever happened to sell the cheapest GPU that quarter.