Explore benchmarks
Filter by model, hardware, workload and quantization. Every score links to the comparisons and experiments behind it.
Top Engines
| # | Engine | Speed score out of 100 |
|---|---|---|
| 1 | uzu†, an approximate comparison, see details | 99 /100 Score |
| ||
| 2 | MLX LM†, an approximate comparison, see details | 96 /100 Score |
| ||
| 3 | Rapid-MLX†, an approximate comparison, see details | 96 /100 Score |
| ||
| 4 | Magnitude†, an approximate comparison, see details | 92 /100 Score |
| ||
| 5 | Ollama†, an approximate comparison, see details | 89 /100 Score |
| ||
| 6 | llama.cpp†, an approximate comparison, see details | 87 /100 Score |
| ||
| 7 | LiteRT-LM†, an approximate comparison, see details | 28 /100 Score |
| ||
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
The score combines counted comparisons. Select one comparison to see its measured values.
† 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. The experiments use different effective KV cache precision: litert-lm default, f16, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64. The experiments use different effective KV cache precision: f16, model-dtype.
† 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. The experiments use different effective KV cache precision: litert-lm default, f16, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64.
† 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. The experiments use different effective KV cache precision: litert-lm default, f16, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64. The experiments use different effective KV cache precision: f16, model-dtype.
† llama.cpp: approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default. The experiments use different effective KV cache precision: litert-lm default, f16, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64. The experiments use different effective KV cache precision: f16, model-dtype.
† 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.
Coverage
| Desktop | Mobile | |||
|---|---|---|---|---|
| M5 Max 64 GiB macOS 27 Apple M5 Max | OnePlus 13 16 GB Android 16 Qualcomm Snapdragon 8 Elite (SM8750) | Pixel 8 8 GB Android 16 Google Tensor G3 | iPhone 12 4 GB iOS 27 Apple A14 Bionic | |
| llama.cpp | 878/8 out of 100, in 8 of 8 comparisons | 1004/4 out of 100, in 4 of 4 comparisons | 893/3 out of 100, in 3 of 3 comparisons | 763/3 out of 100, in 3 of 3 comparisons |
| LiteRT-LM | 284/8 out of 100, in 4 of 8 comparisons | 784/4 out of 100, in 4 of 4 comparisons | 983/3 out of 100, in 3 of 3 comparisons | Not measured on this device. |
| uzu | 994/8 out of 100, in 4 of 8 comparisons | Not measured on this device. | Not measured on this device. | 1003/3 out of 100, in 3 of 3 comparisons |
| Magnitude | 924/8 out of 100, in 4 of 8 comparisons | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| MLX LM | 968/8 out of 100, in 8 of 8 comparisons | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| Ollama | 898/8 out of 100, in 8 of 8 comparisons | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| Rapid-MLX | 966/8 out of 100, in 6 of 8 comparisons | Not measured on this device. | Not measured on this device. | Not measured on this device. |
Speed
| Rank | Engine | Prefill |
|---|---|---|
| 1 | uzu†, approximate comparison, see details | 13,385 tok/s |
| 2 | Magnitude†, approximate comparison, see details | 10,049 tok/s |
| 3 | Ollama†, approximate comparison, see details | 8,827 tok/s |
| 4 | MLX LM†, approximate comparison, see details | 8,486 tok/s |
| 5 | Rapid-MLX†, approximate comparison, see details | 7,775 tok/s |
| 6 | llama.cpp†, approximate comparison, see details | 7,575 tok/s |
| 7 | LiteRT-LM†, approximate comparison, see details | 1,015 tok/s |
details
LFM2.5 1.2B Instruct on M5 Max · 64 GiB · 8-bit · Archive 256 input / 128 output · macOS 27
Medians with the model warmed; model loading is excluded. Each table uses one matched comparison.
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.
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.
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.
llama.cpp: † Approximate comparison. The experiments use different effective KV cache precision: litert-lm default, f16, magnitude default, model-dtype, uzu default.
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.
Memory usage by engine
| Rank | Engine | Memory |
|---|---|---|
| 1 | Ollama†, approximate comparison, see details | 1.41 GiB |
| 2 | llama.cpp†, approximate comparison, see details | 1.41 GiB |
| 3 | MLX LM†, approximate comparison, see details | 1.44 GiB |
| 4 | uzu†, approximate comparison, see details | 1.46 GiB |
| 5 | Rapid-MLX†, approximate comparison, see details | 1.69 GiB |
| 6 | LiteRT-LM†, approximate comparison, see details | 1.73 GiB |
| 7 | Magnitude†, approximate comparison, see details | 1.90 GiB |
details
LFM2.5 1.2B Instruct on M5 Max · 64 GiB · 8-bit · Archive 256 input / 128 output · macOS 27
Memory ranks require comparable, calibrated measurements. Unavailable measurements remain unranked.
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.
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.
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.
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.
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 comparison
Each engine’s Speed score in each matched comparison, out of 100, or why it has none. How rankings work