Skip to content

SilentSwarm Roadmap

SilentSwarm's north star is a self-hostable network for pooling GPU capacity — train and serve models bigger than any single machine, across the mismatched hardware you already have. We're building toward that in three phases, and we're honest about which one works today.

The three phases

Phase Trust model Status
1. Trusted cluster Machines you own or fully control ✅ Works today
2. Federated pools Friends / orgs sharing capacity 🚧 In progress
3. Open marketplace Strangers, metered & isolated 🔭 Planned

Phase 1 (today). A leader plus one or more fellows you control. Pipeline parallelism across mismatched GPUs and machines, NAT-friendly relay transport with zero network config, one-curl fellow bootstrap, train and serve on the same topology.

Phase 2 (federating). Multiple people contribute nodes to shared pools with accounts, per-pool authorization, and usage accounting. Much of the groundwork is in: user accounts, pools with role-based quotas, node credentials, operator auth, and CSRF are implemented. Remaining: metering/accounting and per-pool resource guarantees.

Phase 3 (marketplace). Strangers rent compute to strangers. This needs sandboxing (fellows run untrusted work in isolation), fault tolerance (a node can vanish mid-job), and verification/reputation. Not there yet — and we won't claim it until it is.

Technical priorities (ordered by leverage)

  1. N-stage pipeline. Generalize the currently 2-stage pipeline to arbitrary node counts — the single biggest architectural limit for larger swarms.
  2. Fault tolerance. Stage checkpoint/resume and re-placement so a dead fellow doesn't mean a dead job.
  3. Sandboxing. Container isolation for fellow-executed work (GPU passthrough). The prerequisite for Phase 3.
  4. Honest benchmarks. Published tokens/sec and boundary-bandwidth numbers: relay vs P2P, with/without the learned-bottleneck compression.
  5. Accounting ledger. GPU-seconds / tokens-served metering per fellow — the seed of the marketplace, immediately useful for shared clusters.

What differentiates SilentSwarm

Versus Petals, Exo, GPUStack, and Ray: the learned-bottleneck compression that makes WAN pipeline parallelism economical, NAT-friendly relay transport with zero network config, train and serve on the same topology, and the one-curl fellow bootstrap.

Have an opinion on the ordering, or want to own one of these? Write to info@nexpatch.ai — this roadmap is meant to be shaped in the open. How changes are made is in the contributor guide.