<!-- agent-source-envelope -->
# Artifact: Huge Preset Decode Throughput | PTLLM

> The native Mac runtime reached high local decode throughput on the Huge preset, showing the serving path is viable for local eval loops.

Canonical: https://posttrainllm.com/artifacts/huge-decode-throughput
<!-- /agent-source-envelope -->

Mac runtime benchmark

Report artifact

2026-06-06

This is a runtime artifact, not a model-quality claim. Its value is operational: local eval loops need throughput, stable serving, and cheap repeated generation.

MLX, decode, Mac, throughput

## Headline Numbers

Evidence class applies to each value: measured = retained evaluation or
runtime result; historical = recorded result not freshly reproduced;
derived = calculated or decided from records; observed = current artifact
state; not-measured = an explicit evidence gap.

**Huge decode:** 696 tok/s

96M/Huge preset, ctx 1024

Evidence: historical

**Mega pilot:** 293 tok/s

960M pilot

Evidence: historical

**Warm TTFT p99:** 5.8ms

reported runtime metric

Evidence: historical

## Competitive Context

| System | Metric | Score | Size / Class | Comparable? | Readout |
| --- | --- | --- | --- | --- | --- |
| posttrainllm Huge preset | decode throughput | 696 tok/s | 96M | Direct | Local runtime baseline for cheap repeated eval/smoke loops. |
| posttrainllm Mega pilot | decode throughput | 293 tok/s | 960M | Direct | Shows the throughput drop as local model size approaches specialist scale. |
| External serving stacks | same benchmark | not measured | MLX/llama.cpp/Ollama class | Not comparable | Needs a shared prompt/config/device table before public competitive serving claims. |

Direct rows share this artifact's eval setup. Directional rows are
useful market context but should not be read as leaderboard claims.

## Runtime numbers

| Metric | Value | Use |
| --- | --- | --- |
| Decode throughput | 696 tok/s | Fast local eval/smoke loops |
| Mega pilot throughput | 293 tok/s | Boundary mapping for larger local models |
| Warm TTFT p99 | 5.8ms | Interactive serving viability |

## Release Blockers

### Preset-specific

The headline number is not a blanket claim for all HF models or specialists.

**Unblock:** Attach latency/RAM/tok-s numbers to each future specialist artifact.

## Evidence

- [README headline metrics](https://github.com/PostTrainLLM/posttrainllm#headline-results)

- [Performance docs](https://posttrainllm.com/docs/performance/performance)

## Next Release Action

Use this as the baseline expectation for future artifact performance tables.
