Rankings

Explore benchmarks

Filter by model, hardware, workload and quantization. Every score links to the comparisons and experiments behind it.

Top Engines

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

† Rapid-MLX: approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

† MLX LM: approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

† uzu: approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mirai-m, where most use Q4_K_M.

† Magnitude: approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, 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: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.

† llama.cpp: approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.

ExecuTorch: Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.

LiteRT-LM: Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.

Coverage

Speed score per device, out of 100: each device ranked on its own.
DesktopMobile
M5 Max 64 GiB macOS 27 Apple M5 MaxOnePlus 13 16 GB Android 16 Qualcomm Snapdragon 8 Elite (SM8750)Pixel 8 8 GB Android 16 Google Tensor G3iPhone 12 4 GB iOS 27 Apple A14 Bionic
llama.cpp62 out of 100, in 1 of 1 comparisonsNo rival yet No other engine has a result in the same comparison yet.No rival yet No other engine has a result in the same comparison yet.43 out of 100, in 1 of 1 comparisons
uzu90 out of 100, in 1 of 1 comparisonsNot measured on this device.Not measured on this device.100 out of 100, in 1 of 1 comparisons
Magnitude82 out of 100, in 1 of 1 comparisonsNot measured on this device.Not measured on this device.Not measured on this device.
MLX LM100 out of 100, in 1 of 1 comparisonsNot measured on this device.Not measured on this device.Not measured on this device.
Ollama67 out of 100, in 1 of 1 comparisonsNot measured on this device.Not measured on this device.Not measured on this device.
Rapid-MLX100 out of 100, in 1 of 1 comparisonsNot measured on this device.Not measured on this device.Not measured on this device.
ExecuTorchNot counted Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.Not counted Its output check failed: 5/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.Not counted Its output check failed: 5/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.Not counted Its output check failed: 5/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.
LiteRT-LMNot counted Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.Not counted Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.Not counted Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128.Not measured on this device.

Speed

TTFT ms

Generation tok/s

Speed, LFM2.5 350M on M5 Max · 64 GiB, 4-bit · Archive 256 input / 128 output · macOS 27
RankEngineDecode and TTFT
1Rapid-MLX, approximate comparison, see details Rapid-MLX contributors1,108 tok/s 24.5 ms TTFT
2MLX LM, approximate comparison, see details Apple (ml-explore)1,102 tok/s 23.2 ms TTFT
3uzu, approximate comparison, see details Mirai1,001 tok/s 67.9 ms TTFT
4Magnitude, approximate comparison, see details Magnitude911.9 tok/s 18.7 ms TTFT
5Ollama, approximate comparison, see details Ollama742.8 tok/s 21.4 ms TTFT
6llama.cpp, approximate comparison, see details ggml-org684.9 tok/s 23.2 ms TTFT
Not rankedExecuTorch PyTorch (Meta)Not counted
Not rankedLiteRT-LM Google AI EdgeNot counted
details

LFM2.5 350M on M5 Max · 64 GiB · 4-bit · Archive 256 input / 128 output · macOS 27

Medians with the model warmed; model loading is excluded. Each table uses one matched comparison.

Rapid-MLX: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

MLX LM: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

uzu: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mirai-m, where most use Q4_K_M.

Magnitude: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, 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: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.

llama.cpp: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.

ExecuTorch: Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128. † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-4w-g64, where most use Q4_K_M.

LiteRT-LM: Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128. † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: litert-int4, where most use Q4_K_M.

Memory usage by engine

details

LFM2.5 350M on M5 Max · 64 GiB · 4-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: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.

llama.cpp: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.

uzu: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mirai-m, where most use Q4_K_M.

MLX LM: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

Rapid-MLX: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

Magnitude: † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, 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.

ExecuTorch: Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128. † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: mlx-4w-g64, where most use Q4_K_M.

LiteRT-LM: Its output check failed: 6/8 of the checks the model passes at BF16; needs 7 and the mandatory ones (needle-start, needle-end). Left out, failed by every BF16 configuration: arithmetic, sort, word, continuation-128. † Approximate comparison. The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default. Its weights are 4-bit in another scheme: litert-int4, where most use Q4_K_M.

Every comparison

Comparisons in this view: each engine's score in each comparison, out of 100, or why it has none.
ComparisonRapid-MLXMLX LMuzuMagnitudeOllamallama.cppExecuTorchLiteRT-LM
LFM2.5 350M on M5 Max · 64 GiB 4-bit · Archive 256 input / 128 output · macOS 27 100, approximate100, approximate90, approximate82, approximate67, approximate62, approximateNot countedNot counted

Each engine’s Speed score in each matched comparison, out of 100, or why it has none. How rankings work