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.
Phase A return packet
Outcome: Phase A ready for hardware qualification. The macOS candidate is real-model validation pending, not fully validated.
Source revisions actually selected
| Repository | Revision | Role |
|---|---|---|
| Strata (integration checkout) | branch strata/rebuild-2026-10, HEAD e0951ce243511597eb243018191f194af9d5f8bc |
the product |
| Strata (pristine reference) | 296b59d2 (main) |
unchanged baseline |
| DeepSeek Harness | 57190202 |
agent loop and durable sessions |
| HyperCore-Tune | 4f9ab7fa |
evaluated for tuning; decision D07 |
| ggml (llama.cpp) | 3cf03257 |
engine host sources |
| metal-cpp | 4e6c9a1a… SHA-256 4df3c078…85a3a4 verified |
Metal host bindings |
No upstream drift: every pin above is the revision that was reviewed.
Artifacts
DMG apps/desktop/src-tauri/target/release/bundle/dmg/Strata_0.1.0_aarch64.dmg
sha256 e18c2f4ca60e0142177dc67789b63469f4365bca46860c284173d55b603766fd
hdiutil verify VALID; packaged self-test 8/8 from inside the mounted image
Terminal release/artifacts/strata-terminal-0.1.44.1-darwin-arm64.tar.gz
bytes 42480892
sha256 fafa0076fd30f9fa3d3f1250d8afbf39225bddb80e07a3d818bae2b01e2814d8
self-contained: bundled CPython 3.12.15, verified with no usable system Python
Engine build-metal/strata
sha256 aef25be339b60ad56d34c192847bd59e4351779eed6a1de9166f81836041ee3f
built from source on this Mac; Metal kernel parity 17/17 with one known defect
Install and run, migration and recovery, and every evidence path are in
docs/work/evidence/artifact.txt and docs/work/QUALIFICATION-KIT.md.
Acceptance standing
Every one of the 67 requirement rows of the handoff's acceptance checklist has an explicit
result. python packaging/acceptance_matrix.py --check fails if a row is missing.
| Result | Count |
|---|---|
| passed | 46 |
| failed | 0 |
| pending-hardware | ? |
| pending-release | ? |
| unsupported | 3 |
| not-implemented | ? |
No requirement row is recorded as failed. Denying a defect by deferring it to another machine is
the one thing this matrix refuses to do; failed means the implementation is wrong and is never
written as pending-hardware.
Exact limitations
- Real-model inference has not run and cannot on this machine.
include/strata/core/geometry.hppfixes the compiled shape (n_embd 2560, 48 layers, 512 experts, vocab 248320), so no small or synthetic model is possible; the smallest pack is 37.6 GB mapped plus a 28.8 GB n-gram table, anddocs/MAC.mdmeasures 32 GB as the smallest tested Mac against this host's 16 GiB with ~21 GiB free. Everything downstream of inference is exercised against the mock and scripted workers, and every such result is labelled as a harness test rather than a performance result. - A tuning sweep has never run against a real model. The measurement harness, job, report, apply and rollback are all exercised, but against a stub engine whose numbers mean nothing.
- The Windows and Linux packages are recipes, not builds. The NSIS and AppImage targets are configured and validate; neither has been produced. Windows SYCL and Intel-Mac Metal are refused as unsupported combinations rather than attempted.
- The Windows owner transport is not a named pipe with an ACL. The Python standard library cannot create one, so Windows uses a loopback-only TCP listener with a 256-bit control token. That is weaker than the checklist's stated condition and is recorded as such rather than claimed.
- No signing, notarization, SBOM or font/asset rights evidence. The artifacts are explicitly labelled unsigned internal development candidates. This blocks public release and does not block this milestone.
- One open engine defect:
moe_mm_parityfails withmtl: no function mm_gu_42_16_64_16_2. - Two acceptance rows are not implemented: the offline bundle format (P09) and connector operation assertions (U09). Both are real gaps, both are outside Phase A's implementation scope, and both are listed rather than hidden.
What a Phase B run must do
docs/work/QUALIFICATION-KIT.md has executable procedures, exact prerequisites and expected
assertions for every machine-dependent check, covering installation, shared UI/CLI sessions, real
inference and tools, cancellation and recovery, Time-skip, tuning baseline/search/apply/rollback,
offline reuse, updates and uninstall retention. Section 10 is the one to read first: it is how to
tell a real pass from a vacuous one, written because three checks in this build passed for reasons
unrelated to the feature they named.