ModelBest

What ModelBest is

ModelBest, known in Chinese as 面壁智能, is an artificial intelligence company registered as 北京面壁智能科技有限责任公司.^[1] Its work centres on large models, with a strong emphasis on small models built to run on phones and other local devices rather than only in data centres.

Origins

The company was founded in 2022, when Liu Zhiyuan, an associate professor at Tsinghua University, led its founding.^[1] ModelBest lists Liu as a co-founder and chief scientist.^[1]

The MiniCPM series

According to the company's site, ModelBest has released the MiniCPM edge model series since February 2024.^[2] The site describes the series as offering 8B sparse and 0.5B lightweight variants, with inference up to 220 times faster and a 128K long context.^[2]

Those labels point at the trade-offs edge models are built around. A lightweight variant is small enough to fit on modest hardware; a sparse model activates only part of its parameters for each input, which cuts computation; a longer context window means the model can take in more text at once before it has to stop.

Multimodal models on device

The 35B-A3B variant of MiniCPM-o/V is described as the first open-source MoE multimodal model that runs on device, needing just 3GB of memory on iOS and Android.^[2] MoE, or mixture of experts, splits a network into specialised sub-networks and routes each input to a subset of them, so that only part of the model does work at any given step. That is one way to keep a large model inside a phone's memory budget.

ModelBest says MiniCPM-V 4.7 was released together with OpenBMB on 13 October, in 35B-A3B and 1B versions.^[2]

Speech and training tools

VoxCPM is described as an end-to-end, tokenizer-free speech synthesis model covering 30 languages and 9 Chinese dialects under the Apache 2.0 licence.^[2] Tokenizer-free means the system works without a separate stage that first turns audio or text into discrete tokens; Apache 2.0 is a permissive open-source licence that allows reuse and modification.

The company says its ForgeTrain framework was written autonomously by AI, matched NVIDIA Megatron-LM in eight hours and surpassed it within 1.5 to 2 days, improving measured H100 MFU by about 8% to 10%.^[2] MFU, model FLOPs utilisation, measures how much of a chip's peak arithmetic throughput is actually used during training.

Embodied intelligence

On 19 July, ModelBest released the MiniCPM-Robot series at WAIC, its first embodied intelligence models.^[2] Embodied models are those meant to act through a physical body, such as a robot, rather than only answer queries in text.

Where it sits

Taken together, the releases described on the company's site cover an edge language-model series, multimodal models, speech synthesis, a training framework and robot-oriented models.^[2] The recurring theme is efficiency: smaller size, faster inference and a smaller memory footprint, so that a model can run where the user is rather than only on a server.

Two details show how that theme carries into the company's public positioning. MiniCPM-V 4.7 was released together with OpenBMB, and VoxCPM was published under the Apache 2.0 licence, both of which put work out where others can build on it.^[2] ForgeTrain, meanwhile, is presented not as a product but as evidence about how the company trains models — a framework that, by its account, reached and then passed an established NVIDIA training stack on measured hardware efficiency.^[2] The MiniCPM-Robot series extends the same small-model approach toward embodied use, where memory and latency budgets are tighter still.^[2]