Mference 0.1.0 — tested source release

Mference 0.1.0 — tested source release

Native Swift + Metal inference on Apple Silicon, with a streaming model
installer, CLI, loopback OpenAI-compatible server and Open WebUI launcher.

This release contains source, not a signed Mac application or model
weights. First use requires a build and a separate checkpoint download.

Download and run

Download these three assets together:

• mference-0.1.0.tar.gz
• mference-0.1.0.json (commit/tree and archive provenance)
• mference-0.1.0-SHA256SUMS

On an Apple Silicon Mac with macOS 15+, Swift 6.1+ / Xcode 16.3+ and uv:

shasum -a 256 -c mference-0.1.0-SHA256SUMS
tar -xzf mference-0.1.0.tar.gz
cd mference-0.1.0
./mference-ui.sh doctor
swift build -c release --force-resolved-versions
./mference-ui.sh install gemma4
./mference-ui.sh

Gemma 4 is the first-use recommendation (~14 GB installed). Allow additional
disk for dependencies/builds; first download, build and verification take time.
Interrupted installs resume with ./mference-ui.sh install gemma4 --resume.
For CLI chat use --chat or --messages-file; --prompt is deliberately raw
completion input and does not add a chat template. The CLI/server do not need
uv or Python.

Both services are loopback-only by default and have no application-level
authentication/TLS. Do not expose them through a public proxy or tunnel.
Checkpoint licenses are separate; no weights or user chat data are bundled.

Included

• Source-faithful Swift-Qwen identity, chat/reasoning policy, tool argument
  correction and UI reasoning-history handling, as an optional checkpoint.
• Batched production prefill improvements, bounded expert execution, and
  truthful execution/memory diagnostics with per-family evidence.
• Flash-Next routing/target-state/rollback work and the GLM INT8 ragged-input
  bounds correction.
• Verified model switching, history reuse, streaming, cancellation recovery,
  and tool-result continuation. Expected cancellations no longer log as 500s.
• Reproducible source packaging, exact-commit manifest and SHA-256 checksums.

Validation and limits

Release commit: 4ff63ff555c137d924f024cd3fad8e0b486959f2.
Archive SHA-256: 1299845e54c6c070c174036b253f2032537eaf25b19a0d22ac6c7346d7928498
(15,898,066 bytes). All three CI-produced files match the locally tested bundle
byte-for-byte.

Release gate workflow
validates macOS 15 / Swift 6.1 and macOS 26, then extracts the distribution and
rebuilds/retests it before uploading assets. Local extracted-source validation
on an M3 Ultra / 256 GiB / macOS 26.3 additionally exercises real Gemma and
Qwen 3.6 CLI/server/browser behavior.

The final CI extracted-archive job also passed: release build 251.16s, test
build 229.29s, and 1,297 tests / 233 suites in 677.072s with two known
issues. Executable help, launcher/adapter/packaging tests, archive checksums
and all 74 distributed Markdown files passed before artifact upload.

Local extracted-source release build: 98.57s. Serial package suite:
1,297 tests / 233 suites passed in 280.004s, with two known issues.
Launcher/adapter/packaging checks and all 74 distributed Markdown files pass.
The browser gate covers saved history, prefix reuse, model switching, visible
streaming, Stop and recovery; the API gate checks valid tool JSON and a
simulated tool-result continuation. These timings are validation footers,
not community-protocol speed benchmarks.

A fresh, isolated installation of pinned Open WebUI 0.11.3 also passed doctor,
new-database startup and a real browser → Gemma → answer check. It installed
263 Python packages without modifying the existing UI environment. First UI
startup took about 52 seconds; the build and model download are additional.

Two pre-existing test known issues are disclosed: the optional absent
Flash-Next toy checkpoint and the disabled native draft's row-zero precision
comparison. Neither is represented as a passed native-MTP qualification.
Ordinary CI does not run environment-gated installed-model tests.

Hosted results: macOS 15.7.9 / Swift 6.1.2 passed 1,296 tests in 925.675s;
macOS 26.6.2 / Swift 6.3.3 passed 1,297 in 724.105s. The count difference is an
SDK-conditional Metal 4 declaration test. Swift 6.1 reports three known-issue
entries because it records the same absent fixture twice; these are the same
two underlying conditions, not an additional hidden failure. These hosted jobs
do not establish a physical minimum-RAM recommendation.

GitGuardian's PR check passed. GitHub dependency alerts are disabled for this
repository, so dependency-vulnerability coverage is not certified by this gate.
This release is not a comprehensive security audit; keep both services local.

This source release does not complete the optimization roadmap. Native
Flash-Next MTP remains disabled; Swift-Qwen is optional, not promoted over base
Qwen; GLM completed-answer speed and fresh physical smaller-Mac/context
qualification remain open. No new benchmark or universal quality claim is made.

See the versioned
installation guide,
checkpoint limits,
prefill matrix
and prior candidate evidence.
Documents frozen in the tag describe the pre-publication candidate state; this
release page records publication of that exact tested source.