Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++
1.9 KB
Raw Blame

Native feature parity — 14 October 2026

Consumer flow: choose a compatible model, optionally read pictures, download/load, chat. Native installation uses Strata's own installer, including shared shards, MTP, packing, vision and resumable downloads. Expert mode exposes device placement, native configuration and parallel instances. Settings > Engine contains tools, measurements and diagnostics. No model downloads, native updates or calibration are run merely by opening the UI.

Implementation checklist:

  • Native control adapter: official MCP installer/catalog, isolated native engine instances, measured telemetry, jobs and cancellation.
  • Model catalog and first-run choices from installed Strata's tables, with honest hardware fit and existing-file reuse.
  • Image understanding: optional encoder setup and image payloads; generation stays a separate option.
  • Responses gateway and three protocol connection templates.
  • Native parameter editor with validation, original configuration preservation and backups.
  • MCP manager and chat tool opt-in.
  • Download/startup recovery and engine logs.
  • Measured benchmark, calibration, native updates and engine rollback.
  • All-model Monitor speed table and per-instance details.
  • Focused fixture tests, TypeScript/build, live UI evidence and handover.

The design host remains available for existing preview conversations. Native operations are explicit actions and use a configured local Strata checkout. Real hardware speed and memory allocation are shown only when reported. Backend absence produces a clear setup path, never a simulated successful operation.

Verification: 149 tests passed; TypeScript and production build passed. Browser checks cover consumer loading, expert devices, multiple-instance Monitor, MCP saving and benchmark results. See docs/native-engine-handover.md for contracts and actual GPU verification limits. The official native checkout remained unmodified.