The Moon and Sixpence Recap: Craft Beer Meets Hardcore Talks as 100+ Makers Ignite the On-Device Agent Community
On June 13, 2026, at 18:30, a hardcore tech gathering exclusively for makers — accompanied by craft beer foam — kicked off at the "Knowledge Distillation" space of AGI Bar in Beiyang AI Town, Xuhui District, Shanghai. The event was jointly hosted by ModelBest × OpenBMB open-source community × AGI Bar. More than 100 developers, hardware makers, industry leaders, and tech geeks gathered on site and, amid the collision of craft beer and hardcore talks, completed a deep conversation about the future of on-device agents.
More than a party: deep polishing before the "moon landing"
The event's theme was "The Moon and Sixpence." As Jing Chenzhe, event initiator and head of ecosystem at the OpenBMB community, put it:
"We believe AGI is the moon in the sky, but everything we do every day is a very small footprint compared with that goal. It is precisely because of the moon's distant goal that everything we do today carries meaning."
The event broke the traditional "sit in rows and listen to a lecture" format, combining a Bring Your Own Hardware mechanism with an immersive Moon Fragment Treasure Hunt. Participants came with development boards, smart glasses, and all kinds of "weapons," navigating the immersive AGI Bar scene to complete challenges and collect moon fragments and limited-edition "sixpence" souvenir coins.
The "Coastal Echo" phone booth: capturing makers' honest words in sound
As the event's signature interactive installation, the "Coastal Echo" phone booth (VoxPhone ☎️) became the most popular spot of the night. It runs VoxCPM2, a new-generation speech model self-developed by OpenBMB x ModelBest x THUHCSI, and with a retro red handset as its entry point, it invited every maker present to leave their own echo.
➤ VoxCPM2
🔗 https://github.com/OpenBMB/VoxCPM
During the evening peak, the "Coastal Echo" phone booth collected 46 authentic voice recordings from developers. Unlike conventional industry surveys, this effective feedback struck directly at core issues such as pain points of large models, on-device model needs, and shortcomings of AI hardware, sketching the most vivid real picture of on-device AI. The makers' honest words mostly centered on two topics:
On-device needs return to "fast, private, specialized." Frontline developers no longer blindly pursue the "omniscient" large-parameter models; they urgently need scenario-specific tools that are "fast" (low latency), "private" (privacy protection), and "specialized" (task-specific optimization). This forms a clear complement to the "generalization" pursued by cloud large models.
The breakthrough point for hardware experience lies in being "proactive." Several hardware makers pointed out sharply that the pain point of current AI hardware is "passive triggering." A true qualitative change in interaction requires devices to leap from "waiting for instructions" to "all-day proactive perception and triggering."
Notably, in one of the longest recordings collected, a developer with a liberal arts background said bluntly: "Current large model products generally lean toward an engineering mindset, but non-technical roles such as operations and sales are the biggest potential market." This authentic voice from the front line also deeply inspired the developers and speakers present.
MiniCPM5 hands-on test: on-device intelligence lands on the "desktop"
Another hardcore checkpoint on the treasure map was the MiniCPM5 desk pet adventure task at the AGI Bar front desk, where participants could personally meet a smart desk pet 🐱 powered by the MiniCPM5 on-device model and play an interactive adventure game.
➤ MiniCPM5 Desk Pet
🔗 https://github.com/OpenBMB/MiniCPM-Desk-Pet
➤ MiniCPM5
🔗 https://huggingface.co/openbmb/MiniCPM5-1B
Built on an open-source project, this product proved to the geeks on site that on-device models can be not only extremely lightweight but also deliver full emotional value.
Extreme lightweight deployment: only 0.9B parameters, about 2G of memory footprint, and a file size of only about 100 MB. That means an ordinary 8GB-memory device can run it smoothly, freeing on-device intelligence from expensive compute constraints.
Real-time state perception: it can not only chat but also recognize and remind users of a Coding Agent's working state in real time (such as thinking, executing, and tool calling), giving makers refined animated feedback.
User needs and application scenarios emerging: in the dedicated MiniCPM5 interviews accompanying the event, many participants from enterprises and the individual developer community showed strong interest in MiniCPM5-1B and the desk pet form. They generally recognized its performance in on-device text processing, information extraction, code generation, and other tasks, and hoped to put it to work in content production, office assistance, and local agents.
Speaker highlights: diverse paths from "sixpence" to "the moon"
Nine industry experts from academia, industry, and the open-source community delivered hardcore talks that night, piecing together the future panorama of on-device agents from underlying ecosystems to vertical applications.
Theoretical paradigm: the geometric descent of intelligence density
In the opening, Jing Chenzhe, initiator of the OpenBMB community and head of ecosystem, laid out the future paradigm of on-device large models and the community's vision. He shared the "Densing Law," already published in Nature Machine Intelligence: roughly every 3.3 months, the intelligence density of a large model with the same parameters doubles, and the number of parameters needed for the same intelligence halves. This frontier theory means that high-quality, low-cost "terminal brains" are accelerating their descent, and intelligence will become ubiquitous across every industry in the physical world, reshaping the hardware industry chain. AI-era intelligent hardware must first solve three core pain points on the device side: privacy, real-time performance, and reliability. Today, OpenBMB community models have been downloaded tens of millions of times, and GitHub Stars have surpassed 140,000, providing a solid open-source foundation for the on-device explosion.
Startup signpost: hardware is the ultimate entry point for human-AI interaction
Sun Donglai, founder of Mozhao, who transitioned from software to hardware, brought a highly charged set of practical insights. He said that large models are currently devouring the software application layer at a frantic pace, and software startups in 2025 will fall into extreme involution; rather than standing still, it is better to jump out of the red ocean and do hardware, seizing the ultimate entry point of human-computer interaction. With a "don't reinvent the wheel" software mindset, he hand-built a demo in just over half a month that racked up 200 million views across the internet, efficiently validating market demand. Its core product, the "Mozhao Silencing Mask," aims to solve the privacy and social costs of voice input in public spaces through micro-innovation. Before brain-computer interfaces become reality, body-worn AI devices need an elegant interaction medium, and on-device models will play the key "router" role responsible for front-end privacy and preprocessing.
Following this entry-point thinking, Xu Xingjing, product lead of ModelBest's Pawbie, dug into the product secrets of AI companion hardware. She said AI hardware startups must first clarify the target users and market, then combine AI capabilities. Through deep research, the Soulsoft sub-brand targeted pet substitutes and geek communities, focusing on emotional value; the core of building a deep emotional connection lies in "mutual understanding" and "memory." The product is equipped with audio-visual dual-modal input and also runs a local AI model that needs no internet connection, absolutely safeguarding user privacy.
Xiao Shubo, head of entrepreneur relations at Shenzhen InnoX Academy, shared from the perspective of hard-tech startup incubation the systematic methodology by which Professor Li Zexiang successfully incubated 272 hard-tech companies over more than two decades. Relying on the Greater Bay Area's complete industrial chain and policy advantages, the academy is committed to providing young people with full-chain empowerment — from market demand entry to product definition and rapid mass production — through core courses such as "mentor project-based accompaniment" and "first-principles workshops." Under the wave of "redoing hardware with AI," hardware startups should not pursue long development cycles but should persist in small steps and rapid iteration, precisely solving real pain points driven by Design Thinking. By integrating a supply chain resource library, the academy effectively lowers barriers for startup teams in mold opening and mass production, allowing ideas to truly become hard-tech products with market competitiveness.
Ecosystem foundation: the "compute powerhouse" of open-source AI acceleration
Software development engineer Pan Jiahang delivered a full-stack hardware and software talk on the open-source AI acceleration ecosystem. As one of the top three chip players, AMD is using its fully open-source low-level runtime software ROCm to benchmark and break the traditional closed-source ecosystem, enabling developers to efficiently run AI workloads on graphics cards, integrated GPUs, and other devices. Pan highlighted the AI-MAX 395, an on-device chip built for AIPC, which with top-tier specs such as a 16-core CPU and 128G of memory shows extreme acceleration advantages in local deployment of on-device models and vertical domain tasks.
Hands-on practice: the best benchmark is in the physical world
Zhang Xiaowei, founder of Yuanhang Technology and a serial entrepreneur, brought the most industry-deep perspective. He said many technologies are perfect in demos but fall apart at real industrial sites, because the best context is on site, not in the prompt; only the site is the one benchmark that tests on-device AI performance without leaking the answers.
The site is the assembly workshop of agents and the context window of the physical world; an agent's answer must integrate business rules and IoT data on site to close the loop. To this end, he shared PhoneClaw, an open-source project that gained 1,000+ GitHub Stars within a month and runs entirely locally on a phone, and offered an on-device agent industry self-check list covering dimensions such as "high-frequency repetition, high labor cost, and device silos," giving makers on site a clear guide to commercial implementation.
Maker ecosystem: going deep in narrow scenarios and delivering a crushing productivity blow
AI tech bloggers Xiao Bai and Lin Yi, with 5 million followers across platforms, shared from the unique perspective of all-media creators their recent "Vibe Coding" practice that has swept the geek circle. The frequent updates of large models and Codex this year have let ordinary people cross the code divide and easily build interesting hardcore projects, declaring that "code, like text and video, is becoming a form of self-expression everyone can master."
Xiao Bai shared that after half a month of intensive Codex use, he madly built more than 20 creative projects and successfully launched "VibeCoding Radar," an aggregation site for novice programmers. The project attracted tens of thousands of users within just one week of launch, with daily active users exceeding 2,000. Self-media content products have a strong agile iteration sense of the internet and product instinct, and in the future can collaborate deeply with traditional product people to uncover hits. At the same time, he advocated that AI-native applications should adopt a "build while publishing" traffic-leverage strategy, rapidly accumulating IP and user feedback during development and using traffic to leverage product cold start, which resonated strongly with the makers present.
Independent developer Deng Yichun presented TeXada, a math formula input agent project. The project runs entirely locally on a Mac and uses a small MiniCPM model of about 1B parameters to go deep in a narrow domain, supporting natural language and handwritten image input and outputting perfect LaTeX notes, successfully solving the industry pain points of slow cloud formula rendering and network dependency.
Niu Niujiang, a master's student in accounting at Xiamen University, shared a hardcore project that uses AI to tackle financial report OCR extraction. He chose a small MiniCPM model for header normalization and table splitting, combined with a code engine to verify cross-statement consistency, and can verify an annual report in 7 minutes on an ordinary computer, cutting the 100% human workload caused by a 1% error rate in financial report OCR by a full 95%.
This is an endpoint, and even more a starting point. We are incredibly excited to see chip vendors, hardware makers, model researchers, and individual developers break down barriers and truly come together here. This deep collision in the on-device AI ecosystem sent every maker home with their own inspiration and spark 🔥.
These electronic echoes captured by VoxCPM2 will serve as precious moon-landing logs on the journey to AGI, preserved forever at AGI Bar. The craft beer gathering may be over, but the makers' moon journey continues...
In July, at WAIC (World Artificial Intelligence Conference) in Shanghai, we look forward to seeing you again 👋!