Peer‑to‑peer agents for quant ML

Quantitative finance is the business of training models on price, volume, and other market data to make forecasts or trading decisions. Hyperspace is pitching a way to run those experiments across a peer-to-peer network of autonomous agents instead of a single lab cluster. (oxford-man.ox.ac.uk, hyper.space) The project’s public materials describe Hyperspace as a live blockchain and peer-to-peer network for AI agents, with testnet version 1.3.0 shipped on April 8, 2026. Its website says agents can discover one another, verify work, and settle sub-cent payments through payment channels while running on desktop, browser, or command-line clients. (hyper.space) Hyperspace’s agent network page says each node gets an autonomous “agent brain” that runs a 30-second loop: observe the network, plan work, execute experiments, journal results, and evolve. The same page says nodes can share findings through GossipSub messaging, store rankings in conflict-free replicated data types, and run training jobs in Python on graphics processors or in TypeScript in the browser. (agents.hyper.space) The training method at the center of the pitch is Low-Rank Adaptation, or LoRA, a way to fine-tune a model by freezing most of its original weights and training a much smaller set of adapter weights. Microsoft’s original LoRA paper said that cuts the number of trainable parameters sharply, and PyTorch and Nvidia both document LoRA as a lower-memory way to adapt large language models. (microsoft.com, docs.pytorch.org, docs.nvidia.com) Hyperspace’s public GitHub repository frames the system as “experimental” and says autonomous agents collaboratively train models, share experiments through peer-to-peer gossip, and push results into a living research repo. The repository shows about 1,300 stars, 140 forks, and 120 commits as of April 12, 2026. (github.com, github.com) The same repository says the network publishes hourly snapshots of research state and that anyone can join from a browser or command line. Hyperspace’s GitHub organization also says its broader inference network has more than 2 million nodes and 3.6 million downloads, though those figures come from the company’s own materials. (github.com, github.com) That matters for finance research because market models are expensive to retrain and easy to overfit, especially when data changes quickly. A distributed setup lets many agents test small variations in parallel, while LoRA-style adapters let them swap in new domain tweaks without retraining a full base model each time. (link.springer.com, microsoft.com) Hyperspace is also trying to solve the trust problem that comes with letting software agents delegate work to one another. A March 15, 2026 paper by Varun Mathur for Hyperspace AI describes “AgentRank,” a PageRank-style system that weights agent reputation by verified computation, uptime, and recency rather than simple self-assertion. (agentrank.hyper.space) The company’s own pages present the network as infrastructure for “distributed machine learning research” and “multi-agent systems at scale,” not as a live trading venue or audited financial product. The public evidence so far is a testnet, documentation, research notes, and GitHub activity rather than third-party performance records for investment use. (hyper.space, agents.hyper.space, github.com) So the story here is less about a proven trading engine than about a new lab setup: many small agents, shared over a peer-to-peer network, training and ranking one another’s work in public. Hyperspace says that system is already live in experimental form; whether it produces durable finance results is the part still being tested. (github.com, hyper.space, agentrank.hyper.space)