ChengyiX/freetoken-macos-deployment

FreeToken macOS MLX Reproducibility Evidence

Machine-readable, redacted evidence for a native macOS Apple-Silicon deployment of the FreeToken deployment interface, produced by the companion chengyixu/freetoken-macos-deployment kit and validated against its published JSON Schema.

Dataset description

Experiment ID	freetoken-macos-mlx-20260822-001 (immutable)
Recorded	2026-08-22T19:32:23Z
Outcome	Pass — healthy server accepted max_tokens=1, returned non-empty content (HTTP 200/200, 0.291 s)
Schema	schema/one-token-experiment.schema.json (v1)
Network exposure	127.0.0.1 (loopback) only
Redaction	No credentials, tokens, or private endpoint data

Contents

File	Description
data/freetoken-macos-mlx-20260822-001.json	Passing one-token OpenAI-compatible API experiment
data/macos-mlx-implementation-2026-08-22.md	Methods, pinned dependencies, model revision, and interpretation boundary

Provenance and pins

Component	Pinned value
Upstream software	FreeToken v0.1.2, commit 0ab982f10905fa775962a4eddcb44caa50065251
Compatibility implementation	freetoken-macos-compat-mlx
Model	mlx-community/Qwen3.5-4B-MLX-4bit, revision 32f3e8ecf65426fc3306969496342d504bfa13f3
Runtime	mlx-lm==0.31.3, mlx==0.32.1, mlx-metal==0.32.1
Host	macOS 27.0, Apple Silicon arm64
Sampling	temperature=0, max_tokens=1, streaming disabled

Intended purpose

This dataset exists to make the macOS compatibility claim falsifiable: anyone can diff the recorded request/response against the schema and the pinned configuration, then re-run the documented protocol on their own Apple Silicon host to confirm or refute the result.

Out-of-scope uses

This is a single one-token smoke-test record. It must not be used as:

• evidence that FreeToken's CUDA kernels, cache/offload scheduler, or bandwidth-adaptive MoE method execute on Apple Silicon;
• a throughput, latency-distribution, memory, or cost benchmark (one token is a control-plane smoke test, not a performance measurement);
• evidence of model quality — the recorded completion is not an evaluation result;
• a substitute for the upstream paper's experiments (arXiv:2608.16157), which remain independently unreplicated here.

Limitations and generalization

Results come from one host (macOS 27.0, one Apple-Silicon generation), one quantized checkpoint, one MLX runtime version, and one trial. They do not generalize to other Macs, model families, or runtimes. Unified-memory pressure and warm-cache state are uncontrolled. See data/macos-mlx-implementation-2026-08-22.md for the full interpretation boundary and the companion repo's threats-to-validity document.

Reproduction

git clone https://github.com/chengyixu/freetoken-macos-deployment.git
cd freetoken-macos-deployment
scripts/bootstrap_macos_compat.sh   # pinned venv + checkpoint (~2.85 GiB, first run)
scripts/run_macos_compat.sh         # loopback-only server on 127.0.0.1:1919
# second terminal:
python3 scripts/verify_one_token.py --base-url http://127.0.0.1:1919 \
  --model mlx-community/Qwen3.5-4B-MLX-4bit \
  --runtime freetoken-macos-compat-mlx \
  --runtime-backend 'Apple MLX via mlx-lm 0.31.3, offline local checkpoint' \
  --model-revision 32f3e8ecf65426fc3306969496342d504bfa13f3 \
  --experiment-id freetoken-macos-mlx-YYYYMMDD-001 \
  --output reports/freetoken-macos-mlx-YYYYMMDD-001.json
python3 scripts/validate_experiment.py reports/freetoken-macos-mlx-YYYYMMDD-001.json

Citation

Please cite both this dataset and the upstream system:

@article{yang2026freetoken,
  title   = {FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution},
  author  = {Yang, Shuo and Fan, Xiaoze and Pan, Melissa and Xi, Haocheng and Wang, Zhe and Sun, Shanlin and Keutzer, Kurt and Han, Song and Zaharia, Matei and Xu, Chenfeng and Stoica, Ion},
  journal = {arXiv preprint arXiv:2608.16157},
  year    = {2026}
}

Kit citation metadata: CITATION.cff (Xu, Chengyi, FreeToken macOS deployment kit, v0.3.1, 2026).