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 | Ollama†, an approximate comparison, see details | 98 /100 Score |
| ||
| 2 | llama.cpp†, an approximate comparison, see details | 94 /100 Score |
| ||
| Not ranked in this view | ||
| Not ranked | MLX LM | Too few results |
| Not ranked | Rapid-MLX | Too few results |
| Not ranked | LiteRT-LM | No rival yet |
| Not ranked | Magnitude | Unsupported |
| Not ranked | uzu | Unsupported |
| Not ranked | ZML | Unsupported |
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.
† Ollama: approximate comparison. 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.
† llama.cpp: approximate comparison. 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.
MLX LM: Results in 1 of 2 comparisons; a rank needs 2.
Rapid-MLX: Results in 1 of 2 comparisons; a rank needs 2.
LiteRT-LM: No other engine has a result in the same comparison yet.
Magnitude: Unsupported in 3 comparisons. Magnitude's chat API renders LFM2.5 2.6B's own template, which always opens a thinking block, so it streams reasoning and can't run the closed-think prompts (measured in qualification).
uzu: Unsupported in 2 comparisons. uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification).
ZML: Unsupported in 1 comparison. zml can't run the frozen prompts: Error: LLMD streamed reasoning before its answer; the harness can't time it apart from the answer (measured in qualification).
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 | 942/2 out of 100, in 2 of 2 comparisons | No rival yet No other engine has a result in the same comparison yet. | Not counted Partial: 5 of 6 planned trials were valid. | Unsupported Unsupported in 1 comparison. Its output check failed: 7/9 of the checks the model passes at BF16; needs 8 and the mandatory ones (needle-start, continuation-128). Left out, failed by every BF16 configuration: color, sort, needle-end. |
| Ollama | 982/2 out of 100, in 2 of 2 comparisons | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| LiteRT-LM | No 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. | No rival yet No other engine has a result in the same comparison yet. | Not measured on this device. |
| uzu | Unsupported Unsupported in 2 comparisons. uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification). | Not measured on this device. | Not measured on this device. | Unsupported Unsupported in 2 comparisons. uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification). |
| Magnitude | Unsupported Unsupported in 3 comparisons. Magnitude's chat API renders LFM2.5 2.6B's own template, which always opens a thinking block, so it streams reasoning and can't run the closed-think prompts (measured in qualification). | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| MLX LM | 1/2Not ranked. Results in 1 of 2 comparisons; a rank needs 2. | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| Rapid-MLX | 1/2Not ranked. Results in 1 of 2 comparisons; a rank needs 2. | Not measured on this device. | Not measured on this device. | Not measured on this device. |
| ZML | Unsupported Unsupported in 1 comparison. zml can't run the frozen prompts: Error: LLMD streamed reasoning before its answer; the harness can't time it apart from the answer (measured in qualification). | Not measured on this device. | Not measured on this device. | Not measured on this device. |
Speed
| Rank | Engine | Decode and TTFT |
|---|---|---|
| 1 | MLX LM†, approximate comparison, see details | 163.1 tok/s 53.2 ms TTFT |
| 2 | Rapid-MLX†, approximate comparison, see details | 162.6 tok/s 61.2 ms TTFT |
| 3 | Ollama†, approximate comparison, see details | 155.1 tok/s 52.7 ms TTFT |
| 4 | llama.cpp†, approximate comparison, see details | 148.8 tok/s 62.8 ms TTFT |
| Not ranked | LiteRT-LM | Not counted |
| Not ranked | Magnitude | Unsupported |
| Not ranked | uzu | Unsupported |
details
LFM2.5 2.6B 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.
MLX LM: † Approximate comparison. 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.
Rapid-MLX: † Approximate comparison. 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.
Ollama: † Approximate comparison. 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.
llama.cpp: † Approximate comparison. 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.
LiteRT-LM: Its output check failed: 7/9 of the checks the model passes at BF16; needs 8 and the mandatory ones (needle-start, continuation-128). Left out, failed by every BF16 configuration: color, sort, needle-end. † Approximate comparison. 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: Magnitude's chat API renders LFM2.5 2.6B's own template, which always opens a thinking block, so it streams reasoning and can't run the closed-think prompts (measured in qualification).
uzu: uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification).
Memory usage by engine
| Rank | Engine | Memory |
|---|---|---|
| 1 | Ollama†, approximate comparison, see details | 2.96 GiB |
| 2 | llama.cpp†, approximate comparison, see details | 2.96 GiB |
| 3 | MLX LM†, approximate comparison, see details | 3.02 GiB |
| 4 | Rapid-MLX†, approximate comparison, see details | 3.28 GiB |
| Not ranked | LiteRT-LM | Not counted |
| Not ranked | Magnitude | Unsupported |
| Not ranked | uzu | Unsupported |
details
LFM2.5 2.6B 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, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64.
llama.cpp: † Approximate comparison. 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.
MLX LM: † Approximate comparison. 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.
Rapid-MLX: † Approximate comparison. 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.
LiteRT-LM: Its output check failed: 7/9 of the checks the model passes at BF16; needs 8 and the mandatory ones (needle-start, continuation-128). Left out, failed by every BF16 configuration: color, sort, needle-end. † Approximate comparison. 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: Magnitude's chat API renders LFM2.5 2.6B's own template, which always opens a thinking block, so it streams reasoning and can't run the closed-think prompts (measured in qualification).
uzu: uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification).
Every comparison
| Comparison | Ollama | llama.cpp | MLX LM | Rapid-MLX | LiteRT-LM | Magnitude | uzu | ZML |
|---|---|---|---|---|---|---|---|---|
| LFM2.5 2.6B on M5 Max · 64 GiB 8-bit · Archive 256 input / 128 output · macOS 27 | 95†, approximate | 91†, approximate | 100†, approximate | 100†, approximate | Not counted | Unsupported | Unsupported | Not run |
| LFM2.5 2.6B on M5 Max · 64 GiB bf16 · Archive 256 input / 128 output · macOS 27 | 100 | 96 | Not run | Not run | Not run | Unsupported | Not run | Unsupported |
| LFM2.5 2.6B on M5 Max · 64 GiB 4-bit · Archive 256 input / 128 output · macOS 27 Not counted: fewer than two scored engines | Not counted | Not counted | Not counted | Not counted | Measured†, approximate | Unsupported | Unsupported | Not run |
Each engine’s Speed score in each matched comparison, out of 100, or why it has none. How rankings work