Comparison Counted
Which engine runs LFM2.5 350M best on M5 Max · 64 GiB?
- Device
- M5 Max · 64 GiB (Apple silicon Macs)
- OS
- macOS 27
- Model
- LFM2.5 350M, 350M parameters
- Weights
- bf16
- Workload
- Archive 256 input / 128 output: 256 tokens in, 128 out
- Decoding
- Greedy, batch 1, context 4,096
- Power
- Plugged in
Ranking
Speed: Decode speed: tokens generated per second.
| # | Engine | Decode and First token |
|---|---|---|
| 1 | MLX LM†, an approximate comparison, see details | Decode: 571.7 tok/s 18.1 ms first token |
| 2 | Rapid-MLX†, an approximate comparison, see details | Decode: 559.0 tok/s 26.2 ms first token |
| 3 | Magnitude†, an approximate comparison, see details | Decode: 524.1 tok/s 19.1 ms first token |
| 4 | Ollama†, an approximate comparison, see details | Decode: 479.5 tok/s 22.3 ms first token |
| 5 | llama.cpp†, an approximate comparison, see details | Decode: 464.2 tok/s 23.4 ms first token |
| 6 | ZML†, an approximate comparison, see details | Decode: 420.2 tok/s 15.9 ms first token |
details
Speed Decode speed: tokens generated per second.
Score: 100 is the best result in every comparison an engine entered; 80 is 80% of the best, on average. Only complete experiments with a passed output check count. How rankings work
† MLX LM: approximate comparison. The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default.
† Rapid-MLX: approximate comparison. The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default.
† Magnitude: approximate comparison. The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. Its phases were timed by the harness clock around the engine API, so they include transport that engine counters leave out.
† Ollama: approximate comparison. The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default.
† llama.cpp: approximate comparison. The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default.
† ZML: approximate comparison. The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. Its phases were timed by the harness clock around the engine API, so they include transport that engine counters leave out.
Every experiment
Medians over valid trials, with the lowest and highest beneath. Eligible experiments are complete and passed their output check. The ranking above selects each engine's configuration for this objective when its required metrics are measured.
| Experiment | Ranking eligibility | Decode tok/s, higher is better | First token ms, lower is better | Memory GiB, lower is better | Prefill tok/s, higher is better | Valid trials |
|---|---|---|---|---|---|---|
| MLX LM 0.32.2 · Metal · BF16 | Eligible, approximately: The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. | 571.7 range565.1–573.2 | 18.1 range9.3–18.3 | 0.94 range0.91–0.94 | 14,490 range14,307–28,852 | 6of 6 planned |
| Rapid-MLX 1.0.2 · Metal · BF16 | Eligible, approximately: The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. | 559.0 range554.2–560.3 | 26.2 range20.2–27.4 | 1.17 range1.14–1.18 | 9,995 range9,554–13,067 | 6of 6 planned |
| Magnitude 0.2.17 · Metal · BF16 | Eligible, approximately: The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. Its phases were timed by the harness clock around the engine API, so they include transport that engine counters leave out. | 524.1 range512.1–526.0 | 19.1 range18.9–19.5 | 1.35 range1.33–1.35 | 13,435 range13,137–13,547 | 6of 6 planned |
| Ollama 0.40.13 · Metal · BF16 | Eligible, approximately: The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. | 479.5 range475.4–483.2 | 22.3 range22.1–22.4 | 0.91 range0.91–0.92 | 12,319 range12,186–12,385 | 6of 6 planned |
| llama.cpp b12110 · Metal · BF16 | Eligible, approximately: The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. | 464.2 range437.9–478.8 | 23.4 range14.2–26.0 | 0.91 range0.91–0.92 | 11,299 range10,150–19,058 | 6of 6 planned |
| ZML 20261106.0 · Metal · BF16 | Eligible, approximately: The experiments use different effective KV cache precision: f16, magnitude default, model-dtype, zml default. Its phases were timed by the harness clock around the engine API, so they include transport that engine counters leave out. | 420.2 range417.3–422.5 | 15.9 range14.9–16.2 | 4.89 range4.85–4.94 | 16,085 range15,849–17,218 | 6of 6 planned |
Every trial, metric by metric
The line runs from an experiment's lowest to its highest valid trial, the dot is the median and each tick is one trial. Every scale starts at zero.
Caveats
- Publisher cards identify the base model but do not always pin the exact export source checkpoint revision.
- Engines may differ in quantization scheme, effective KV cache precision, kernels, and activations.
About this comparison
Worked out from its experiments: every experiment here shares the device, the OS major version, the power state, the base model, the quantization class and the workload. Measured under methodology v0.6. Comparison id glenn-m5-max-64gb--macos-27--lfm2.5-350m--bf16--archive-256-128.