[ToluClassics]

What changed

• pin the exact Llama 3.2 1B and Qwen3 0.6B revisions behind the verified support claims
• record successful public-CLI runtime affine 4-bit/group-64 generation for both checkpoints
• add both models to the reproducibility log for the short-decode and prefill scenarios
• refresh the Phi-3 rows on the merged benchmark implementation after its BOS-token fix
• record the sanitized local environment and exact implementation revision

No package source code or raw benchmark result files are added.

Why

The benchmark runner was verified initially with a quantized Phi-3 checkpoint. This follow-up demonstrates that the same deterministic protocol and public inference APIs work for two actively maintained model families without expanding architecture scope.

Local Apple silicon evidence

All model resolution was strict offline against already-cached immutable snapshots.

Model	Scenario	Median TTFT	Median prefill	Median decode	Peak MLX memory
Qwen3 0.6B	short-decode-128	24.1 ms	2656.7 tok/s	72.4 tok/s	2.95 GiB
Qwen3 0.6B	prefill-512-decode-64	98.6 ms	5192.9 tok/s	65.1 tok/s	4.10 GiB
Llama 3.2 1B	short-decode-128	28.6 ms	2236.3 tok/s	51.9 tok/s	4.77 GiB
Llama 3.2 1B	prefill-512-decode-64	155.9 ms	3284.2 tok/s	50.2 tok/s	5.66 GiB

Each scenario used one warmup and five measured runs. Generated-token checksums matched across every measured run.

Additional inference checks

• Llama 3.2 1B full-precision bounded generation: passed
• Llama 3.2 1B runtime affine 4-bit/group-64 bounded generation: passed
• Qwen3 0.6B full-precision bounded generation: passed
• Qwen3 0.6B runtime affine 4-bit/group-64 bounded generation: passed

Validation

• exact PR Add reproducible MLX benchmark protocol #33 merge CI: passed
• current environment offline suite: 125 tests, 103 passed, 22 skipped
• minimum compatibility environment offline suite: 125 tests, 103 passed, 22 skipped
• ruff check .
• ruff format --check .
• six protocol v1 result validations
• wheel and source distribution build
• twine check on both artifacts