Skip to content

Rampage 0.2.0 — Model Fabric

Pre-release
Pre-release

Choose a tag to compare

@Chris-ObtuseAI Chris-ObtuseAI released this 01 Aug 09:50
961353c

Rampage 0.2.0 — Model Fabric

Rampage 0.2 turns a group of trusted machines into an explicit strategy surface for local AI. The
desktop now separates the two goals people usually blur together: fitting the largest possible
model and producing the fastest possible response.

What changed

  • Maximum Model plans the largest qualified, compatibility-matched aggregate memory placement.
  • Speed Boost accepts distributed tensor peers only when signed link measurements predict a
    real gain; slow links fall back to the best whole-model host.
  • Throughput, Efficiency, and Autonomous Balanced expose replica, smallest-fit, and
    proposal-only adaptive planning lanes.
  • Signed agent offers can advertise bounded Ollama, Exo/MLX, or vLLM/Ray runtime profiles through a
    strict manifest and digest contract.
  • The controller, CLI, Rust SDK, TypeScript SDK, Python SDK, and native desktop all share the same
    versioned model-session planning contract.
  • The native Tauri shell now has a role-aware Windows tray, close-to-tray behavior, Start with
    Windows, quiet background launch, emergency stop, explicit Quit, and deterministic sidecar
    cleanup.
  • MSI and NSIS packages create an automatic desktop launcher plus a Start-menu Rampage Shell whose
    session-local PATH makes the authenticated CLI immediately usable without mutating system PATH.

The boundary that matters

The 0.2 planner is real and read-only. It never turns a capacity estimate into execution authority.
Cross-host model launch and the proposed loopback inference gateway remain disabled until a backend
passes correctness, isolation, topology, streaming-failure, cleanup, and measured-benefit gates.
That distinction is the foundation for making Rampage ambitious without making it reckless.

Verification

The release candidate passes 44 Rust tests, strict all-target Clippy, 4 desktop tests and production
build, 5 TypeScript SDK tests, 13 intelligence tests, 5 Python SDK tests, authenticated direct-QUIC
mesh and artifact proofs, a real Ollama generation receipt, native owner/worker lifecycle probes,
and a silent NSIS install/shortcut/uninstall smoke.

The Windows artifacts are still unsigned. See release evidence before
distribution.