AFM 0.9.16 · release evidence

Qwen 3.8 release qualification

PublishedAugust 15, 2026Qualified targetmlx-community/Qwen3.8-27B-4bit

Release-candidate qualification for Qwen 3.8 27B covering the Swift suite, live model behavior, CLI integration, comprehensive inference, and agentic Promptfoo scenarios.

Qualification results

Swift suite420 passed

One opt-in real-checkpoint test skipped.

Live assertions118 / 118

All live model assertions passed.

Comprehensive182 / 182

No-AI-judge suite.

CLI integration15 / 15

Pipe and command integration coverage.

Promptfoo347 / 388

89.43% across agentic and structured-output cases.

Observed decode30–38 tok/s

Measured during this qualification run.

Interpretation

Known caveats stay visible.

The 41 Promptfoo misses were semantic tool-selection or assertion mismatches; no server crash, transport failure, or parser/runtime failure was observed.
Some CLI pipe tests accept any nonempty fallback response and do not strictly distinguish explicit -s input from redirected stdin in every case.

Bundle inventory

43 evidence files.

Live assertion report in HTML and JSONL
CLI integration HTML report
Comprehensive no-AI-judge JSONL results
All 38 Promptfoo JSON result files

Canonical evidence bundle

afm-v0.9.16-test-reports.tar.gz

483 KiB compressed archive attached to the AFM 0.9.16 GitHub Release.

Download reports ↗Open the GitHub Release ↗

SHA-256: 09fb01874a54eb6168506baac86950cbbf3c7eb5f61320637505a98ee753f85a
Reference URL: https://maclocal.ai/test-reports/0.9.16