Hyperspace is building a decentralized AI network that uses spare computing resources from personal computers instead of owning GPUs.
2
The new product turns a query into an editable directed acyclic graph of tasks, with web scraping, Python execution, retrieval, and model outputs built into the workflow.
3
Users can change node tasks, add nodes, run Python in a terminal, and inspect outputs while keeping control of the plan.
Summary
Nicolas Schlaepfer introduces Hyperspace's decentralized AI network and a new product built on top of it. The network uses resources contributed by people's personal computers, and its existing AIOS product can route inference to another user's machine. The new application combines prompt engineering, a visual React Flow node editor, Python execution, and web browsing. A query is converted into a directed acyclic graph in JSON, where each node receives a task and produces an output. Users can edit the plan, add nodes, run generated Python, and inspect results in a terminal. Hyperspace also describes a web-scraping primitive based on Puppeteer and Beautiful Soup, plus a virtual file system. The demo uses Qwen 2 Instruct for reasoning and Llama 3 70B for summarization. Schlaepfer presents these features as groundwork for future agent behavior, including memory, planning, code generation, and execution.
Hyperspace runs inference through a community network of personal computers
Nicolas Schlaepfer describes Hyperspace as a decentralized AI network with no GPUs owned by the company. People contribute resources from their personal computers to the network. Its existing AIOS product is available for Windows and Mac and uses llama.cpp for inference. A user can download the product and have a chat experience that gets inference from another participant, such as someone in Belgium. The company is building around shared computing resources rather than a centrally hosted service.
A mixture of specialist models is central to Hyperspace's AI approach
Schlaepfer says Hyperspace does not believe one large closed-source model is the answer for the best AI experience. He argues that a mixture of several strong expert models can provide a better result. This idea leads into the new Hyperspace product, which is designed to run on the company's decentralized network. The product combines prompt engineering with a visual React Flow interface, Python execution, and retrieval-like web browsing.
The planning model turns a query into an editable task graph
The first model Hyperspace built for the product is a fine-tuned model for agentic planning. A user submits a query and receives a directed acyclic graph in JSON. The graph is a methodical plan derived from the query. Each node represents a task in the plan, and the resulting structure can be used by the product's orchestration system. The goal is to give users a concrete plan that can be inspected and changed rather than hiding the entire process inside one model call.
Web pages are converted into text that language models can read
Hyperspace is also building an in-house web-scraping primitive for language models. It uses Puppeteer and Beautiful Soup to scrape websites. The scraped HTML is converted into Markdown, which Schlaepfer describes as easier for language models to digest. This browsing capability is one part of the wider product, alongside planning, visual workflow editing, and code execution.
The node editor lets power users change the generated workflow
The product has a node editor and a terminal intended as a power tool for power users. The editor is based on the directed acyclic graph orchestration model. Users can change each node in the React Flow interface to fit their needs. In the demo, the system streams nodes into the editor from the Hyper Engine V3 planning model. Users can edit the title, task description, and expected output, and they can add as many nodes as they want.
Each node combines the overall goal with its local task and earlier results
When the workflow runs, every node produces an output. Each node creates a query from the overall goal, its local goal, and what happened in the previous node. Hyperspace uses a reasoning model and a summarization model in this process. Schlaepfer names Qwen 2 Instruct for reasoning and Llama 3 70B for summarization. Using both gives the workflow a set of synthesized answers from different stages of the plan.
Python outputs can be executed directly in a terminal
Some nodes produce Python. When that happens, the user can run the code and inspect the result in a terminal. The terminal opens automatically with the output. This connects the visual task graph to actual computation and gives users a way to execute parts of a generated plan instead of only reading model text.
The product is assembling primitives for future agent behavior
Schlaepfer says Hyperspace wants to provide the groundwork for agentic behavior through a set of basic capabilities. The demo includes planning, Python execution, and code generation, along with a virtual file system that changes the working directory. He says the team is building the primitives an agent would need, including memory. The product was announced as an early release that would become available through WeightList later that week.
"We want to provide the groundwork for agentic behavior in the future by providing these core primitives."04:42
Who should watch
You are building agent workflows and want users to inspect and edit the plan instead of accepting a hidden chain of model calls.
Your application needs to combine language models with web scraping, Python execution, a file system, and a visual task graph.
You are interested in peer-to-peer inference and in routing model work through computing resources contributed by personal devices.
