Explore this comparison’s ranking

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
4-bit
Workload
Archive 256 input / 128 output: 256 tokens in, 128 out
Decoding
Greedy, batch 1, context 4,096
Power
Plugged in

Speed: Decode speed: tokens generated per second.

Which engine runs LFM2.5 350M best on M5 Max · 64 GiB? Ranked by Speed.
#EngineDecode and First token
1Rapid-MLX, an approximate comparison, see details Rapid-MLX contributors Decode: 1,108 tok/s 24.5 ms first token
2MLX LM, an approximate comparison, see details Apple (ml-explore) Decode: 1,102 tok/s 23.2 ms first token
3uzu, an approximate comparison, see details Mirai Decode: 1,001 tok/s 67.9 ms first token
4Magnitude, an approximate comparison, see details Magnitude Decode: 911.9 tok/s 18.7 ms first token
5Ollama, an approximate comparison, see details Ollama Decode: 742.8 tok/s 21.4 ms first token
6llama.cpp, an approximate comparison, see details ggml-org Decode: 684.9 tok/s 23.2 ms first token
Not ranked in this view
Not rankedExecuTorch PyTorch (Meta) 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 rankedLiteRT-LM Google AI Edge 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.
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.

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.

Every experiment in this comparison
ExperimentRanking eligibilityDecode tok/s, higher is betterFirst token ms, lower is betterMemory GiB, lower is betterPrefill tok/s, higher is betterValid trials
Rapid-MLX 1.0.2 · Metal · mlx-affine-4bit-g64 Eligible, approximately: 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.1,108 range1,100–1,11024.5 range23.9–25.60.72 range0.68–0.7410,658 range10,202–10,9496of 6 planned
MLX LM 0.32.2 · Metal · mlx-affine-4bit-g64 Eligible, approximately: 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.1,102 range1,097–1,10523.2 range23.2–23.30.50 range0.50–0.5011,258 range11,206–11,2726of 6 planned
uzu 0.6.17 · Metal · mirai-m Eligible, approximately: 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.1,001 range996.6–1,00367.9 range67.2–68.70.48 range0.48–0.5319,884 range19,381–20,8256of 6 planned
Magnitude 0.2.17 · Metal · Q4_K_M Eligible, approximately: 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.911.9 range904.9–915.918.7 range18.4–25.01.05 range1.05–1.0613,679 range10,229–13,9226of 6 planned
Ollama 0.40.13 · Metal · Q4_K_M Eligible, approximately: The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.742.8 range734.2–745.721.4 range21.3–21.60.46 range0.46–0.4712,795 range12,754–12,8706of 6 planned
llama.cpp b12110 · Metal · Q4_K_M Eligible, approximately: The experiments use different effective KV cache precision: bf16, fp32, litert-lm default, f16, magnitude default, model-dtype, uzu default.684.9 range667.8–718.623.2 range13.0–23.70.46 range0.46–0.4711,475 range11,176–21,4516of 6 planned
ExecuTorch 1.6.0 · Metal · mlx-4w-g64 Complete 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-128.786.4 range779.1–788.823.4 range23.1–23.5Not measured11,130 range10,667–11,1306of 6 planned
ExecuTorch 1.6.0 · CPU · 8da4w Complete 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-128.335.5 range334.2–340.585.8 range85.7–86.7Not measured2,977 range2,943–3,0126of 6 planned
LiteRT-LM 0.19.1 · CPU · litert-int4 Complete 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-128.183.5 range179.4–185.6244 range244–2540.62 range0.62–0.621,118 range1,072–1,1226of 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.

Decode speed tok/s, higher is better
Time to first token ms, lower is better
Peak memory GiB, lower is better
Prefill speed tok/s, higher is better

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--4-bit--archive-256-128.