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 2048 input / 128 output: 2,048 tokens in, 128 out
- Decoding
- Greedy, batch 1, context 4,096
- Power
- Plugged in
Ranking
Balanced: Decode speed, time to first token and peak memory, weighted equally.
| # | Engine | Balanced score out of 100 |
|---|---|---|
| 1 | Magnitude†, an approximate comparison, see details | 82 /100 Balanced score |
| 2 | Ollama†, an approximate comparison, see details | 82 /100 Balanced score |
| 3 | llama.cpp†, an approximate comparison, see details | 81 /100 Balanced score |
| 4 | uzu†, an approximate comparison, see details | 76 /100 Balanced score |
| 5 | MLX LM†, an approximate comparison, see details | 59 /100 Balanced score |
| 6 | Rapid-MLX†, an approximate comparison, see details | 58 /100 Balanced score |
| 7 | LiteRT-LM†, an approximate comparison, see details | 21 /100 Balanced score |
details
Balanced Decode speed, time to first token and peak memory, weighted equally.
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
† 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.
† 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.
† 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.
† 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.
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 |
|---|---|---|---|---|---|---|
| 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. | 747.9 range731.2–756.8 | 44.4 range44.3–44.5 | 1.04 range1.04–1.05 | 46,119 range46,034–46,275 | 6of 6 planned |
| 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. | 599.8 range586.3–603.9 | 61.5 range61.5–61.7 | 0.61 range0.61–0.61 | 34,935 range34,835–35,108 | 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. | 592.1 range568.8–612.1 | 63.4 range61.9–70.3 | 0.61 range0.61–0.61 | 32,938 range29,505–33,763 | 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. | 776.1 range771.4–779.6 | 88.8 range86.9–90.3 | 0.70 range0.70–0.71 | 66,557 range65,117–71,324 | 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. | 702.2 range695.5–703.1 | 57.7 range57.2–57.9 | 2.09 range1.94–2.38 | 35,880 range35,780–36,186 | 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. | 712.2 range711.3–714.6 | 61.9 range59.5–65.3 | 2.05 range1.91–2.35 | 33,458 range31,693–34,790 | 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. | 198.9 range198.3–200.6 | 635 range631–643 | 1.25 range1.25–1.25 | 3,308 range3,268–3,330 | 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-2048-128.