ClystriaClystriaRequest access
Changelog

Everything we've forged.

Clystria is early and moving fast. Here's the trail of sparks.

v0.1.0
2026-07-01
Private beta

First light

  • ▸Gossip-based control plane — no master, no external database.
  • ▸Automatic layer placement across heterogeneous GPUs.
  • ▸Self-healing failover: layers re-forge onto survivors on node loss.
  • ▸OpenAI-compatible /v1 endpoint for every deployment.
  • ▸forge CLI: init, join, deploy, scale, status, top.
v0.0.9
2026-06-12
Preview

The pool takes shape

  • ▸Unified VRAM pool addressable across nodes.
  • ▸Paged KV-cache with cross-node eviction awareness.
  • ▸AWQ / GPTQ / fp4 auto-fit quantization on deploy.
  • ▸Live TUI (clystria top) for per-GPU utilization.
v0.0.6
2026-05-20
Preview

Two machines, one brain

  • ▸Cross-node tensor & pipeline parallelism over RDMA with TCP fallback.
  • ▸Apple Silicon (M-series) agent support.
  • ▸Rolling, zero-downtime model updates.
v0.0.1
2026-04-02
Spark

The forge is lit

  • ▸1.2 MB agent that fingerprints GPUs and joins a network.
  • ▸Single-node model serving as a proof of fire.