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
- 8-bit
- Workload
- Archive 256 input / 128 output: 256 tokens in, 128 out
- Decoding
- Greedy, batch 1, context 4,096
- Power
- Plugged in
Ranking
Memory: Peak memory while generating: the less, the better.
| # | Engine | Memory and Decode |
|---|---|---|
| 1 | Ollama†, an approximate comparison, see details | Memory: 0.60 GiB 630.3 tok/s decode |
| 2 | llama.cpp†, an approximate comparison, see details | Memory: 0.60 GiB 617.2 tok/s decode |
| 3 | uzu†, an approximate comparison, see details | Memory: 0.67 GiB 794.4 tok/s decode |
| 4 | MLX LM†, an approximate comparison, see details | Memory: 0.67 GiB 818.5 tok/s decode |
| 5 | LiteRT-LM†, an approximate comparison, see details | Memory: 0.78 GiB 200.0 tok/s decode |
| 6 | Rapid-MLX†, an approximate comparison, see details | Memory: 0.90 GiB 799.5 tok/s decode |
| 7 | Magnitude†, an approximate comparison, see details | Memory: 1.04 GiB 764.7 tok/s decode |
details
Memory Peak memory while generating: the less, the better.
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
† Ollama: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default.
† llama.cpp: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default.
† uzu: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: mirai-l, where most use Q8_0.
† MLX LM: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: mlx-affine-8bit-g64, where most use Q8_0.
† LiteRT-LM: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: litert-int8, where most use Q8_0.
† Rapid-MLX: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: mlx-affine-8bit-g64, where most use Q8_0.
† Magnitude: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu 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 |
|---|---|---|---|---|---|---|
| Ollama 0.40.13 · Metal · Q8_0 | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. | 630.3 range624.6–645.1 | 21.4 range21.1–21.5 | 0.60 range0.60–0.61 | 12,907 range12,819–13,117 | 6of 6 planned |
| llama.cpp b12110 · Metal · Q8_0 | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. | 617.2 range590.2–655.2 | 22.6 range21.8–23.0 | 0.60 range0.60–0.61 | 11,913 range11,629–12,350 | 6of 6 planned |
| uzu 0.6.17 · Metal · mirai-l | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: mirai-l, where most use Q8_0. | 794.4 range790.4–795.5 | 69.2 range66.9–71.4 | 0.67 range0.67–0.71 | 19,693 range18,755–20,523 | 6of 6 planned |
| MLX LM 0.32.2 · Metal · mlx-affine-8bit-g64 | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: mlx-affine-8bit-g64, where most use Q8_0. | 818.5 range814.7–820.5 | 23.6 range23.4–23.9 | 0.67 range0.65–0.68 | 11,082 range10,953–11,168 | 6of 6 planned |
| LiteRT-LM 0.19.1 · CPU · litert-int8 | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: litert-int8, where most use Q8_0. | 200.0 range197.7–201.0 | 136 range136–139 | 0.78 range0.78–0.78 | 2,103 range2,050–2,108 | 6of 6 planned |
| Rapid-MLX 1.0.2 · Metal · mlx-affine-8bit-g64 | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 8-bit in another scheme: mlx-affine-8bit-g64, where most use Q8_0. | 799.5 range794.3–820.0 | 29.2 range25.1–31.3 | 0.90 range0.88–0.92 | 8,949 range8,338–10,390 | 6of 6 planned |
| Magnitude 0.2.17 · Metal · Q8_0 | Eligible, approximately: The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. Its phases were timed by the harness clock around the engine API, so they include transport that engine counters leave out. | 764.7 range759.1–768.9 | 18.0 range17.8–19.1 | 1.04 range1.03–1.04 | 14,188 range13,437–14,349 | 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--8-bit--archive-256-128.