In the AI PC Era, True "Token Freedom" Is on the Edge: Edge Computing Engine Company AgiBot Secures Tens of Millions of RMB in Funding

Wange AI's team carries a few defining labels: Gen Z, PhD-heavy, heads-down on tech.

CEO Guanbo Wang checks all three boxes. He's currently pursuing his PhD at Tsinghua University's Department of Computer Science, and he's a serial entrepreneur born after 2000.

The team is remarkably young — around 20 people, nearly 90% of them Gen Z. Most hold master's or doctoral degrees from schools like Tsinghua and Peking University, with others coming from Amazon, OpenAI, and ByteDance.

Recently, Wange AI closed two consecutive funding rounds — an angel round and an angel+ round totaling tens of millions of RMB, with 5Y Capital and FreeS Fund participating, and Yuanhe Capital serving as exclusive financial advisor. The proceeds will go toward product R&D and market expansion.

In the past, moving compute to the cloud was practically mandatory. But with the explosion of agent capabilities from Claude Code, Codex, and OpenClaw, token demand has grown exponentially.

Guanbo Wang puts it bluntly: none of the inference engines on the market are well-suited for edge devices. Most existing engines focus on speed while ignoring massive memory consumption.

On the edge, chip makers rarely offer more than 32GB of RAM. Too much memory limits use cases. So for hardware vendors, the goal is clear: make their chips infer faster and run larger models within existing memory constraints, without adding hardware costs.

Wange AI's answer is cPilot, an edge compute engine, paired with Amis, an intelligent platform — a solution designed to give users cheap, usable tokens:

On cost, it lets small-memory machines run large models, slashing hardware requirements for deployment. On performance, it targets large models on the edge — not small ones — delivering local deployment that actually solves user needs.

"Some solutions sacrifice speed or accuracy just to force a model to run in low-memory environments. Our edge inference approach is at least 12x faster under the same memory overhead," Wang told Intelligent Emergence.

In 2025, they spent nearly a full year getting their product to work across different chip vendors' hardware — at a time when consumer demand for edge intelligence was still weak.

This year, the viral success of agent tools like OpenClaw opened their eyes to B2C possibilities.

Wang explains that Wange AI's primary customers are currently B2B chip vendors. The company collaborates with them on terminal hardware, installing its edge compute engine and a locally developed product called "Lobster" on AI mini PCs, AI PCs, and AI NAS devices. It also provides an edge compute optimization package with a pre-installed platform for one-click model deployment and API aggregation, meeting consumer demand for local large model deployment.

For now, Wange AI's business model is B2B-first, gradually testing and validating a B2C model through its B2B2C efforts.

The company has entered the delivery phase with multiple hardware partners, expecting tens of thousands of pre-installed units to ship this year. Revenue is projected to exceed 10 million RMB.

No Small Edge Models

Today's large model market is a price war.

Recently, DeepSeek cut API prices for DeepSeek-V4-Pro by 75%. Lei Jun announced price reductions for the MiMo V2.5 series, with cuts as steep as 99%.

The consensus behind this: AI has genuinely entered productive use cases, and users increasingly want good models at low cost.

Wange AI is aligned with this. It targets edge hardware capabilities, letting users run large-parameter models locally — solving the cost problem at its root. Beyond hardware savings, local deployment means zero token costs.

From day one, they decided: no small edge models — the market is too narrow and not general-purpose enough. No post-training — because when cloud models iterate, knowledge gets overwritten.

Following this logic, Wange AI launched cPilot, an edge AI inference engine.

cPilot is a bottom-layer engine, a middleware layer between hardware and software. Through proprietary algorithms, it maximally compresses model memory footprint, unlocking the full potential of underlying hardware.

Typically, a 32GB device can only spare 8–10GB for model inference, locally running roughly a 4B-parameter model.

With cPilot on identical hardware, deployable model sizes jump from 4B to 80B parameters. For one hardware vendor client, using cPilot saved about 2,000 RMB per machine while multiplying the deployable model size.

But local deployment isn't a cure-all. Edge capabilities have limits. Meanwhile, user needs are shifting: as models grow more capable, people no longer blindly chase the most powerful model — they want to call the right model for the right task.

So Wange AI recently launched Amis, an edge intelligence platform that connects to mainstream agent tools and models while also giving users access to cloud compute.

Amis functions as an API aggregation platform and scheduling hub. Users can directly run agent tools like OpenClaw and Hermes, flexibly connecting and switching between models. The platform automatically allocates between cloud and local compute, switching based on task complexity.

The benefit: most user needs are lightweight, high-frequency, token-burning tasks that can be handled locally. Only a small fraction of complex tasks beyond the edge's capabilities need the cloud.

Users don't pay other model providers. They configure models directly on Amis, and through edge-cloud scheduling, most simple tasks complete locally at zero token cost. Only 10–20% of tasks go to the cloud, dramatically compressing costs.

"Our hope is to carve out a solid position in broad consumer应用场景," Wang said. "Amis's ultimate goal is to get users hooked on the platform as an ecosystem habit."

MoE Is Already Sparse,

But There's 10x Room to Improve

Wang believes that if everyone can see a market opportunity, it's not one for startups.

At founding, before MoE (Mixture of Experts) had gained much traction, Wange AI chose to optimize for Dense architectures on the edge.

Many then thought open-source models were too limited — was it too early to bet on edge intelligence?

Wang chose to gamble boldly on the uncertainty of user demand and industry trends.

Three bets: First, model capability — would users settle for models that simply solve their needs, rather than chasing maximum quality? Second, hardware cost, which they decided to make their core moat. Third, whether token consumption would explode.

Anchored on these three points, Wange AI started from optimizing hardware capability and reducing model memory footprint, conducting full-stack optimization across the hardware layer, middleware, and algorithm software.

From the software and algorithm perspective, whether Dense or MoE, only a subset of parameters are activated during inference. Even MoE, which already exploits sparsity, has roughly 10x more sparsity to exploit.

So Wange AI designed a "dynamic sparse activation algorithm" that accurately predicts which parameters to compute and load during inference, drastically reducing actual parameter usage.

From the edge hardware side, three bandwidth bottlenecks constrain overall performance: memory bandwidth, CPU memory access bandwidth, and CPU-GPU interconnect bandwidth. Addressing these, Wange AI built a scheduling system analogous to CUDA (Compute Unified Device Architecture), transforming the hardware layer into an edge large model inference platform and memory management system, with adaptations for different vendors' chips.

In testing, Wang said, they ran a 35B-parameter model on an AMD-powered machine with 27.6GB memory footprint. Under identical hardware with the cPilot engine, that dropped to 4.7GB.

This means users can run models like Qwen3.6 and Gemma 4 — capable of coding and complex tasks — in under 5GB of memory.

The Second Half of AI Is on the Edge

"Edge used to be underrated," Wang told Intelligent Emergence. "But many investors have told us that this year, a consensus is forming across the investment landscape — edge could be the future."

Compared to the explosive growth in agent capabilities and token demand, vendors' token price cuts are a drop in the bucket.

Wange AI wants the edge to become the next computing paradigm, moving users from "renting intelligence" to "owning intelligence."

Long-term, they envision token usage becoming like WiFi today — every device capable of generating tokens locally, moving cloud capabilities entirely to the edge, with each device serving its surrounding network.

Currently, Wange AI remains focused on the middleware between software and hardware. But Wang says this is Phase One.

In the next phase, they may consider developing their own edge AI hardware. "It's not quite the right time for hardware yet," as Wang puts it.

For one, chip technology hasn't converged. Current GPUs work well for training but not efficient inference. Jumping into hardware now would lock in form factors and make later iteration costly. Next-generation chips, like domestic NPUs, may bring a major shift on the chip side.

For another, hardware isn't just about tech and engineering — supply chain capability matters more. "For hardware, we'd need to plan 10 months ahead to build out upstream and downstream supply chains and sales channels," Wang said. "Working with B2B2C customers lets us stake out an ecosystem position early."

In the next phase of edge computing, some application will emerge that can handle the token explosion. Wange AI aims to provide the downstream services enabling such apps. Long-term, they want cPilot and Amis to become the most complete, cross-platform, out-of-the-box solution in the low-memory segment.