Comparison Counted
Which engine runs LFM2.5 2.6B best on M5 Max · 64 GiB?
- Device
- M5 Max · 64 GiB (Apple silicon Macs)
- OS
- macOS 27
- Model
- LFM2.5 2.6B, 2.6B 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
Memory: Peak memory while generating: the less, the better.
| # | Engine | Memory and Decode |
|---|---|---|
| 1 | Ollama†, an approximate comparison, see details | Memory: 2.97 GiB 153.6 tok/s decode |
| 2 | llama.cpp†, an approximate comparison, see details | Memory: 2.98 GiB 142.4 tok/s decode |
| 3 | MLX LM†, an approximate comparison, see details | Memory: 4.01 GiB 160.4 tok/s decode |
| 4 | Rapid-MLX†, an approximate comparison, see details | Memory: 4.14 GiB 160.0 tok/s decode |
| Not ranked in this view | ||
| Not ranked | Magnitude | Unsupported Unsupported in 1 comparison. 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 ranked | uzu | Unsupported Unsupported in 1 comparison. uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification). |
| Not ranked | LiteRT-LM | Not counted 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. |
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, 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.
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, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64. | 153.6 range153.3–153.9 | 283 range283–284 | 2.97 range2.97–2.97 | 7,306 range7,275–7,317 | 6of 6 planned |
| llama.cpp b12110 · Metal · Q8_0 | Eligible, approximately: 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. | 142.4 range140.5–147.3 | 298 range297–299 | 2.98 range2.98–2.98 | 6,885 range6,874–6,908 | 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, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64. | 160.4 range158.8–160.5 | 221 range220–222 | 4.01 range3.92–4.04 | 9,292 range9,277–9,321 | 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, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64. | 160.0 range159.4–160.8 | 242 range240–242 | 4.14 range4.10–4.22 | 8,490 range8,484–8,560 | 6of 6 planned |
| LiteRT-LM 0.19.1 · CPU · litert-int8 | Complete 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. | 60.3 range59.8–60.9 | 2,923 range2,917–2,932 | 4.42 range4.42–4.43 | 711.9 range709.8–713.4 | 6of 6 planned |
| Magnitude 0.2.17 · Metal · Q8_0 | Unsupported 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). | – | – | – | – | 0of 6 planned |
| uzu 0.6.17 · Metal · mirai-l | Unsupported uzu can't run the frozen prompts: Error: uzu counted 23 input tokens, not the frozen prompt's 26 (measured in qualification). | – | – | – | – | 0of 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--lfm2.5-2.6b--8-bit--archive-2048-128.