Model size
—smallerCloud AI · native Arm64 · reproducible CI
Optimization claims
with receipts.
ArmForge CI quantizes real ONNX language models for Arm64, benchmarks them on native hardware, validates numerical quality, and publishes the evidence automatically.
Loading latest benchmark…
Mean latency
—lowerThroughput
—higherOutput cosine
—quality retainedBefore / after
One model. Same inputs.
A fair comparison.
| Metric | FP32 | Arm64 INT8 |
|---|---|---|
| Model size | — | — |
| Mean latency | — | — |
| P95 latency | — | — |
| Throughput | — | — |
The evidence pipeline
Pin
Download an exact model revision and verify its SHA-256 digest.
Forge
Apply per-channel signed INT8 dynamic quantization for Arm64 inference.
Measure
Run isolated FP32 and INT8 workers on a native four-core Arm64 runner.
Prove
Gate numerical quality and publish raw JSON plus a readable report.