Explore this comparison’s ranking

Comparison Counted

Which engine runs LFM2.5 1.2B Instruct best on M5 Max · 64 GiB?

Device
M5 Max · 64 GiB (Apple silicon Macs)
OS
macOS 27
Model
LFM2.5 1.2B Instruct, 1.2B parameters
Weights
4-bit
Workload
Archive 2048 input / 128 output: 2,048 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 1.2B Instruct best on M5 Max · 64 GiB? Ranked by Speed.
#EngineDecode and First token
1uzu, an approximate comparison, see details Mirai Decode: 528.7 tok/s 126 ms first token
2ExecuTorch, an approximate comparison, see details PyTorch (Meta) Decode: 456.8 tok/s 332 ms first token
3Magnitude, an approximate comparison, see details Magnitude Decode: 450.2 tok/s 109 ms first token
4Ollama, an approximate comparison, see details Ollama Decode: 442.1 tok/s 139 ms first token
5llama.cpp, an approximate comparison, see details ggml-org Decode: 436.7 tok/s 145 ms first token
Not ranked in this view
Not rankedLiteRT-LM Google AI Edge Not counted Its output check failed: 10/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.
Not rankedMLX LM Apple (ml-explore) Not counted Its output check failed: 10/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.
Not rankedRapid-MLX Rapid-MLX contributors Not counted Its output check failed: 10/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.
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

† 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.

† ExecuTorch: 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.

† 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
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.528.7 range525.9–529.0126 range124–1300.95 range0.95–0.9531,301 range31,245–31,6116of 6 planned
ExecuTorch 1.6.0 · Metal · mlx-4w-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-4w-g64, where most use Q4_K_M.456.8 range445.6–458.5332 range330–332Not measured6,178 range6,169–6,2066of 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.450.2 range441.1–466.3109 range96.1–1201.56 range1.56–1.5618,730 range17,068–21,3196of 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.442.1 range439.9–445.3139 range138–1410.95 range0.95–0.9515,035 range14,809–15,1996of 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.436.7 range416.3–447.6145 range144–1490.95 range0.95–0.9514,182 range13,797–14,3596of 6 planned
ExecuTorch 1.6.0 · CPU · 8da4w 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: 8da4w, where most use Q4_K_M.121.6 range116.8–128.52,405 range2,357–3,052Not measured851.6 range671.3–869.36of 6 planned
LiteRT-LM 0.19.1 · CPU · litert-int4 Complete Its output check failed: 10/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.104.3 range96.9–114.04,588 range4,527–5,1272.20 range2.19–2.20448.9 range401.5–454.96of 6 planned
LiteRT-LM 0.19.1 · GPU · litert-int4-gpu Complete Its output check failed: 8/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.349.2 range347.9–349.5374 range369–3771.88 range1.88–1.895,799 range5,773–5,8816of 6 planned
MLX LM 0.32.2 · Metal · mlx-affine-4bit-g64 Complete Its output check failed: 10/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.533.3 range531.6–539.697.7 range87.9–98.41.83 range1.82–1.8321,107 range20,972–23,4666of 6 planned
Rapid-MLX 1.0.2 · Metal · mlx-affine-4bit-g64 Complete Its output check failed: 10/12 of the checks the model passes at BF16; needs 11 and the mandatory ones (needle-start, needle-end, continuation-128). Left out, failed by every BF16 configuration: none.538.5 range536.5–541.9105 range103–1061.98 range1.94–1.9919,715 range19,457–20,1046of 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--lfm2.5-1.2b-instruct--4-bit--archive-2048-128.