AFM 0.9.15 · release evidence

DeepSeek V4 release validation

PublishedAugust 12, 2026Qualified targetdeepseek-ai/DeepSeek-V4-Flash-0731-AFM-MLX

Release evidence for the DeepSeek V4 AFM MLX path, including unit tests, live assertions, comprehensive inference, concurrency, prefix reuse, and Promptfoo agent workflows.

Qualification results

Swift suite390 passed

31 Swift Testing suites.

Assertion rerun487 / 487

Three unsupported cases skipped.

Comprehensive182 / 182

No-AI-judge suite.

Promptfoo334 / 388

38 scenario result files; no harness errors.

Concurrent correctness32 / 32

Validated across batch sizes 1, 2, 4, and 8.

Mixed workload32 / 32

Validated across four batch sizes.

Interpretation

Known caveats stay visible.

The initial aggregate assertion run recorded 478/479; the bundled targeted rerun records 487/487 with three skips.
The multi-turn prefix-cache run recorded 39/44 across batch sizes, so the archive preserves both the passing concurrency evidence and the remaining multi-turn misses.

Bundle inventory

300 evidence files.

Swift release-test log
DeepSeek assertion HTML, JSONL, and rerun logs
Comprehensive HTML, JSONL, and per-run server logs
Promptfoo JSON results and server logs
Concurrency, mixed-workload, and multi-turn prefix logs

Canonical evidence bundle

afm-v0.9.15-test-reports.tar.gz

1.64 MiB compressed archive attached to the AFM 0.9.15 GitHub Release.

Download reports ↗Open the GitHub Release ↗

SHA-256: 3a0da609f10cd10fdf69675ecaa80ffcd269bc7da4cb29f23518146fe55130a5
Reference URL: https://maclocal.ai/test-reports/0.9.15