Realtime AI News
ModelBest Releases a Large Supervision Agent Built on Its MiniCPM Base Models
On October 16, ModelBest held a launch event in Beijing and released its Large Supervision Agent, combining risk, internal control, compliance, audit and discipline inspection into one five-in-one supervision capability. Co-created with industry partner Xinghuo Mingzhi and built on the MiniCPM family, it is a flagship result of the ModelBest+ partner programme announced at the World AI Conference in July.
On October 16, ModelBest held a launch event at the Zhongguancun Wenquan Science Park in Beijing and formally released its Large Supervision Agent (大监督智能体).
The product is one of the first major results of the ModelBest+ industry partner programme announced during the World AI Conference in July, bundling risk, internal control, compliance, audit and discipline inspection into a single five-in-one supervision capability.
It was co-created with industry partner Xinghuo Mingzhi (星火明智), and the two companies say they will focus on product co-development, real-world deployment and market expansion. More than a hundred guests attended the launch, including ModelBest co-founder and president Lei Shengtao, Xinghuo Mingzhi founder Zhao Guowei and Li Auto senior vice president Xu Hongbin.
The technical base is ModelBest's MiniCPM family of foundation models. According to the launch materials, the series is known for high knowledge density and strong factual accuracy, which suits high-rigour scenarios, and it has already been deployed at scale in professional fields such as law.
On top of that base, the Large Supervision Agent offers more than 50 supervision skills across nine business domains, along with five agent capabilities that run the full chain from early warning and anomaly detection through case initiation, classification, remediation and closure. The company says around-the-clock patrols turn supervision from people hunting for problems into problems finding people, while cloud-plus-device compute keeps data inside the domain and allows offline use on-device.
It combines ModelBest's large model and agent technology with Xinghuo Mingzhi's more than 20 years of audit and supervision experience, handing procedural supervision work to AI so that supervisors can focus on judgement and deeper investigation, with every step logged, traceable and controllable.
Xinghuo Mingzhi founder Zhao Guowei said supervision is a profession that demands respect, and that the partnership has, for the first time, put the company's two decades of audit experience into an AI-native agent at scale.
ModelBest was founded in 2022 and chose on-device AI from the start; the company describes itself as China's first mover with a full on-device stack. For a field like audit supervision, where data security and process control are paramount, on-device and local deployment is the precondition for AI to reach the front line of real business.
The ModelBest+ model is straightforward: the model and agent framework come from ModelBest, while industry know-how comes from partners. Its AI solutions are already deployed at scale in automotive, phone and embodied robotics scenarios as well as in legal and financial fields, and the Large Supervision Agent is presented as a template for how the programme works in enterprise supervision.
Why it matters
This is a concrete case of an industry agent moving from demo to a professional workflow, pairing two decades of audit experience with large model capability so that procedural supervision goes to AI and judgement stays with people. It also gives the ModelBest+ division of labour between model maker and domain partner a replicable template.
Nearby Updates
All10/16, 22:41
ByteDance Sets Up AI Safety Team and Opens Global Hiring Drive
ByteDance has set up a dedicated AI safety team and opened a global hiring drive for the unit, according to reports carried by multiple outlets. The move signals that one of China's largest AI developers is building out safety capacity as its models and products scale.
10/16, 22:49
Microsoft Finishes Pre-Training MAI-Base-2, a 2.8T-Parameter Model It Says Beats DeepSeek on Efficiency
Microsoft AI said on October 15 that it completed pre-training of MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token, trained on 16,000 Nvidia GB200 GPUs over 38 days. Microsoft claims its recipe is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base, but the model is still in post-training with no release announced.
10/16, 23:17
17-year-old Kimi researcher Chen Guangyu says he has left Moonshot AI
Chen Guangyu, a post-2000 developer known for co-authoring Kimi's Attention Residuals technical report at 17, said on social media on Oct. 15 that he had left Moonshot AI's Kimi team, according to a Southern Metropolis Daily report on Oct. 16. The report does not say why he left or where he is going next.
10/16, 23:21
Nikkei: DeepSeek hit by five-hour outage as V4-Pro and V4.1-Flash go down
Nikkei Asia reports that DeepSeek suffered an outage of around five hours on Thursday, with its V4-Pro and V4.1-Flash models and chat services unusable from 8 a.m. to 1 p.m. The company has not identified a cause, some point to insufficient computing resources, and paid subscribers received credit refunds for services affected between Oct. 8 and Oct. 14.