MiniCPM5

What MiniCPM5 is

MiniCPM5 is a family of small open-weight language models built on dense Transformer architectures. Unlike mixture-of-experts (MoE) designs, which send each token through a selected subset of the model's parameters, every MiniCPM5 model is dense: all of its parameters are used for every token ^[3]. The series is released under the Apache-2.0 licence ^[2]. The team positions it for on-device use, local deployment and other resource-constrained settings rather than for large-scale server-side serving ^[3].

Release timeline

The first model, MiniCPM5-1B, was published on Hugging Face on 21 May 2026 and opened the series ^[1]. MiniCPM5-2B followed on 6 September 2026 ^[2]. The team describes the 2B model as a scaling-up of the same training recipe used for the 1B model, so the two are meant to be read as two points on one line rather than as separate designs ^[3]. That is a common pattern for small-model families: one recipe, several sizes, so that a developer can pick the largest model that still fits the target device.

Why dense at this size

Whether a model is dense or sparse matters a great deal once it leaves the data centre. In a mixture-of-experts model, only part of the network fires for any given token, so the compute per token is smaller than the total parameter count suggests — but the full set of weights still has to be stored somewhere, and the routing step has to be supported by the runtime. A dense model does the opposite: the memory footprint and the arithmetic per token both follow directly from the parameter count, with no routing decisions in between. For a model that is meant to run on a phone, a laptop or a small edge box, that predictability is worth a lot, because memory bandwidth and battery life are the binding constraints rather than raw throughput. Quantisation and local inference runtimes also tend to work most simply on straightforward dense graphs.

What the team claims

The team describes MiniCPM5-2B as 2B-class open-source state of the art, and says it remains competitive with 4B-class models overall ^[3]. Compared with models of a similar size, it claims advantages in coding, mathematics, long-context understanding, tool use and agentic tasks ^[3]. Those are the team's own characterisations of the model; the point worth noting for a reader is the scope of the claim — not just text quality but the skills that make a small model usable inside an application, such as calling tools and carrying out multi-step tasks.

Data and distribution

Training data is released alongside the model as part of the UltraData family, starting with UltraX ^[3]. Publishing the corpus together with the weights makes it possible to see what a small model was trained on, and to reuse that data for other models. Weights ship in BF16, GGUF, MLX and GPTQ, among other formats ^[3]. The spread of formats tracks the spread of places these models are meant to run: GGUF for local CPU-and-GPU inference tools, MLX for Apple silicon, GPTQ for quantised deployment, and BF16 as the reference precision. Adoption has been substantial for a model of this size — the MiniCPM5-2B card recorded more than one million downloads on Hugging Face ^[2].

Where it sits

MiniCPM5 belongs to the branch of open model development concerned with doing more with fewer parameters. The bet behind it is that a well-trained dense model of one or two billion parameters, released openly with its data and in formats that local runtimes accept, is more useful to many developers than a much larger model they can only reach through an API. The 1B and 2B entries are the opening of that series, and later sizes were signalled by the way the first release was framed ^[1].