README.md

---
license: agpl-3.0
tags:
- tribunus
- inference
- benchmarking
- apple-silicon
- mlx
- compute-image
- evidence
---

# Tribunus Benchmarks

Research data from the Tribunus Compute inference engine. Includes benchmark results, backend comparison data, TurboQuant KV cache compression analysis, ComputeImage compile manifests, and evidence receipts.

## Methodology

All benchmarks run on Apple Silicon (M1 Max, 64GB) running macOS 26.5.
- Model: Qwen2.5 0.5B Instruct (NF4 quantized)
- Temperature: 0.0 (greedy)
- Tokens: variable (50-512 per run)
- Warm-up: 3 runs before measurement
- Report: throughput (tok/s), time-to-first-token (TTFT), latency distribution

## Hardware

| Component | Detail |
|-----------|--------|
| SoC | Apple M1 Max |
| RAM | 64GB unified |
| GPU | 32-core Metal GPU |
| ANE | 16-core Neural Engine |
| OS | macOS 26.5 |
| MLX | 0.31.2 (tribunus fork) |

## Datasets

- `benchmarks/qwen2.5-0.5b.json` — Throughput, TTFT, latency distribution
- `benchmarks/backend-comparison.json` — MLX GPU vs Accelerate CPU vs Core ML ANE
- `compute-images/manifest-qwen2.5-0.5b.json` — NF4 quantized, 24 layers, 556 tensors
- `evidence/receipt-samples.json` — Sample receipt data from inference runs

## Results Summary

| Phase | tok/s | Notes |
|-------|-------|-------|
| Baseline | 65 | Custom MLX-heavy runtime with heterogeneous scaffolding |
| Arena/residency | 100-160 | Zero-copy arena, paged residency |
| Speculation active | 180-280 | Speculative decode with ANE draft |
| Full stack | 300+ | Draft model, verifier batching, KV transactions |

## Citation

```
@software{tribunus_compute_2025,
  author = {Julian Torr},
  title = {Tribunus Compute Benchmarks},
  year = {2025},
  url = {https://huggingface.co/datasets/Tribunus-dev/tribunus-benchmarks}
}
```

## License

AGPL-3.0