For the past three years, the AI industry has operated on one assumption: more capability means more cloud. Frontier models live in data centers, and researchers reach them through APIs, subscriptions, and constant internet connections. On September 3, 2026, at a product launch in Shanghai, a Peking University spin-off and HP made the opposite bet.
Metaspace Intelligence, also known as Yuankong Intelligence and rooted in Peking University’s School of Science Intelligence, unveiled Boxer, an edge-side AI model designed to run on ordinary workstations with no internet connection required.
Alongside it came Metaspace AI Work, an offline office platform, and Metaspace AI for Science, an agent built to carry out scientific research workflows end to end. HP confirmed at the same event that it will pre-install the software and models on its Z-series workstations.
The pitch is simple. Cloud-based large models come with problems that keep resurfacing: data security, token costs, and connection stability. Metaspace CEO Pang Dawei argues the edge side is a “physical solution” to all three, and the company spent three years building the agent tooling to make that claim credible.
The demonstration was modest in scope but striking in direction: an off-the-shelf workstation was completing real office and research tasks with no network connection at all.
A Peking University Spin-Off Partners With HP
Metaspace Intelligence was founded in February 2023, and its path to edge AI was not a straight line. The team started with ChatExcel, a consumer chatbot-style spreadsheet tool, then evolved through full-link data agent platforms before settling on edge-side intelligence as its core business. That trajectory matters, because the training data behind its new model comes from millions of real task steps its agents generated along the way.
The HP partnership gives the startup something a small team cannot build alone: distribution. AI Work and the Boxer model will ship pre-installed on HP Z-series machines under a standardized, buyout-style license rather than a per-token subscription. Zhou Xinhong, vice president of China Hewlett-Packard, framed the strategy bluntly: “Every computer in the future may be an AI computer, and the demand for edge-side AI, especially enterprise-level edge-side AI, is about to usher in an explosion.”
It also draws a sharp line between HP and its traditional rivals. While Lenovo and Dell have concentrated on servers, data centers, and cloud-side AI infrastructure, HP is positioning itself purely on the edge, targeting industries where data legally cannot leave the building.
Meet Boxer: 35 Billion Parameters Sized for a Workstation
Boxer is a 35-billion-parameter model, with the version demonstrated at the event labeled Boxer-35B-A3B-v0-0819. It was post-trained on top of an open-source base model, with training data drawn from the real task trajectories generated by Metaspace’s own agent harness. According to the company’s self-built WorkArena evaluation set, it ranked first among models of its size class on real-world task performance.
The performance figures reported by the vendor are the kind usually reserved for cloud deployments:
- Prefill speed of more than 700 tokens per second
- Decoding speed of more than 30 tokens per second
- Less than 8GB of memory in use while running
- Hardware requirement of a consumer-grade PC with roughly 40 TOPS of computing power
How Does a 35B Model Compare to Frontier AI?
Co-founder Dr. Ning Kunpeng claimed the 35B model is more capable than GPT-4o was more than a year ago, a comparison worth treating cautiously since no third-party benchmark backs it yet. Still, the broader trend he pointed to is real: smaller open models keep closing the gap on giant ones. It is the same dynamic behind Qwen 3.8’s 27B open model rivaling far larger proprietary systems, and Ning specifically cited DeepSeek V4 Flash, which activates only 13B parameters per token, as evidence that raw size no longer dictates intelligence density.
Is There an AI Model That Works Offline?
Yes, and the Shanghai launch included a live proof. Dr. Ning ran three real tasks on an HP computer with no internet connection: AI Work classified and archived invoices, flagged anomalies in financial transaction flows within minutes, and handled research tasks that stretch from running complex experiments to drawing charts and drafting papers.
What makes this generation of offline models different from earlier local AI experiments is the agent layer. A local chatbot that answers questions is a novelty; a local system that opens files, executes multi-step workflows, and produces finished deliverables is a productivity tool. For readers who want to explore local options today, open-weight releases such as DeepSeek V3.2’s MIT-licensed local reasoning model already run on consumer hardware, though without the office and research packaging Metaspace is shipping.
What Laptop Can Run AI Locally?
The honest answer in late 2026: any modern AI PC with a neural processing unit in the 40-TOPS class, which describes a growing share of new business laptops and all of HP’s Z-series workstations.
But the more interesting story is what happens to the rest of the installed base.
The Economics of the AI PC Tipping Point
Zhou argued that the economics are approaching a tipping point. PC component prices, from memory to SSDs to processors, have been rising since 2025, and HP’s own data shows the sub-$700 PC market shrinking while GPU-equipped machines above $700 post double-digit revenue growth.
His calculation: once manufacturers integrate large memory pools directly into devices, a local AI workstation will cost roughly what an enterprise spends on cloud AI services over two or three years. At that point, he said, the switch to local is “a natural process.” NVIDIA’s push to pack large memory into chips is accelerating exactly this shift.
With the world shipping around 300 million PCs a year, even a modest share of AI-capable machines represents an enormous base of devices that could run models like Boxer without ever touching a data center.
| Market signal from the Shanghai launch | Figure |
|---|---|
| Global PC shipments per year | ~300 million units |
| PCs priced under $700 | Market shrinking |
| GPU-equipped PCs above $700 | Double-digit revenue growth |
| Enterprise AI deployment spend, 2-3 year forecast | Expected to double |
| Component prices since 2025 (memory, SSD, CPU) | Rising |
From Invoices to Research Papers: What the System Actually Does
The demo tasks hint at two distinct products sharing one local model:
- Metaspace AI Work handles office operations: invoice classification and archiving, anomaly detection in financial transaction records, and similar back-office workflows that normally send sensitive documents to cloud APIs.
- Metaspace AI for Science targets researchers, with the company claiming it can automate the arc of a complex experiment, generate the charts, and draft the paper sections that follow.
The research angle has precedent in the same ecosystem.
Peking University and the YuanKong AI Agent Joint Laboratory open-sourced OpenAI4S, a research agent under the MIT license with more than 30 built-in scientific skills, from protein structure analysis and molecular docking to single-cell analysis and literature retrieval. Boxer is the commercial, hardware-integrated continuation of that work rather than a standalone curiosity.
Why Run AI for Scientific Research Locally?
Cloud AI is powerful, but entire categories of research cannot legally or practically use it.
Pharmaceutical companies, financial institutions, and manufacturers operate under data compliance rules that require knowledge and records to stay on local machines. For researchers handling unpublished results, patient-adjacent data, or proprietary experimental data, sending prompts to an external API is a non-starter.
Three factors make the local case compelling right now:
- Data control. Ning stated that for fully local deployments, data generated during use “will never flow back” to the company. Only enterprise clients that explicitly authorize it contribute training data.
- Predictable cost. A buyout license replaces recurring token bills. As one point of context for why per-token pricing frustrates budgets, Chinese AI models already undercut Western APIs on price, and local inference eliminates the meter entirely.
- Efficiency. A model that runs in under 8GB of memory on a 40-TOPS chip is doing far less environmental work per query than a request routed through a hyperscale data center. Local small models are not automatically greener, but for routine office and analysis tasks they avoid the data center round trip altogether.
What Is the Best AI for Research: Cloud Agents or Local Models?
There is no universal winner, and anyone shopping for the best AI for research should match the tool to the sensitivity and scale of the work.
- Choose cloud models when you need frontier-level reasoning, multimodal analysis, or constant access to the latest weights, and your data can leave the premises.
- Choose local models when data must stay on-site, when internet access is unreliable, when token budgets are unpredictable, or when the tasks are well-scoped office and analysis workflows that a 35B model handles comfortably.
Pang Dawei is candid about the competitive risk: “There is no technology that large manufacturers do not have.” His wager is that cloud vendors’ business models keep them from fully committing to the edge, and that the closed data loops Metaspace built over three years are hard to replicate quickly. If the gap between small and large models keeps narrowing, that wager ages well. If frontier capability pulls away again, edge AI stays a specialist’s tool.
| Factor | Cloud AI | Local (edge) AI |
|---|---|---|
| Data location | Sent to external servers | Stays on-device |
| Cost model | Recurring per-token fees | One-time license |
| Capability ceiling | Frontier models | Mid-size models such as 35B |
| Internet required | Always | No |
| Best fit | Open-ended, heavy reasoning | Compliance-heavy office and research workflows |
Frequently Asked Questions
What are the best offline AI models available right now?
Open-weight models like DeepSeek V3.2 and smaller Qwen releases run locally on capable hardware and are free to download.
Boxer is a new entry aimed specifically at office and research workflows, distributed pre-installed on HP workstations rather than as a public download. The “best” choice depends on whether you need raw capability, structured workflow automation, or strict data isolation.
Does an offline AI model send my data anywhere?
Not by design. Metaspace says fully local deployments never return usage data to the company, and enterprise clients only contribute training data if they explicitly authorize it. The same principle applies to any genuinely local model: if the weights run on your machine and it has no network access, there is nowhere for your data to go.
Is local AI as capable as cloud AI?
Not at the frontier.
Cloud models still lead on the hardest reasoning tasks. Local models like Boxer have become good enough for well-defined office and research workflows, and vendors claim performance comparable to last year’s flagships, but independent benchmarks for Boxer are not yet available.
When will laptops match the cost of cloud AI subscriptions?
HP’s Zhou Xinhong predicts the crossover within two to three years, as component prices rise and large memory pools move on-package. His rule of thumb: when a local AI workstation costs about the same as two or three years of cloud AI service fees, the economics flip permanently toward the edge.
The deeper story from Shanghai is not one model’s benchmark score. It is that the AI PC argument has shifted from “can it run a chatbot” to “can it run your lab.” For researchers deciding how to build AI-era skills, the practical lesson from stories like a high school dropout who became an OpenAI researcher still applies: the tools change fast, but the habit of working alongside them compounds. As reported by 36kr’s interview with the Metaspace and HP executives, the edge-side market is small teams and big iron finding each other. Whether Boxer fulfills its offline promise will be settled by independent testing, but the direction of travel is now hard to deny.
