AskLoci/loci-speed-benchmarks

M1 Pro engine microbenchmarks · September 30, 2026

18 timing observations from six synthetic tests on one Apple M1 Pro Mac with 32 GiB of unified memory. Three Q4_K_M GGUF models were measured with a pinned llama.cpp engine. Each workload has three sequential repetitions.

These are historical engine measurements under substantial host load. They do not measure Loci app speed, model quality, phone performance or memory fit.

Methods and permanent edition · Raw timings and metadata · Measurements CSV · Checksums and provenance

Recorded results

Rates are synthetic model tokens per second. The summary is the arithmetic mean and sample standard deviation of the three individual repetition rates. The standard deviation is not a confidence interval.

Model	Workload	Mean tokens/s	Sample standard deviation	Repetitions
MiniCPM5-2B	pp512	796.31	3.21	3
MiniCPM5-2B	tg128	75.84	0.06	3
Spark-X2.5-1.7B	pp512	1098.80	6.05	3
Spark-X2.5-1.7B	tg128	98.55	0.19	3
Spark-X2.5-4B	pp512	431.35	0.35	3
Spark-X2.5-4B	tg128	44.79	0.04	3

• pp512: Processes 512 synthetic token IDs, with no prior context.
• tg128: Performs 128 single-token decode steps from empty context, without a preceding prompt. It does not generate a sampled natural-language answer.

Machine and method

The recorded machine had an Apple M1 Pro, 10 CPU cores, 16 GPU cores and 32 GiB unified memory. The exact Mac product identifier and operating-system version were not retained.

The runtime was llama.cpp 7fe450e19305, using Metal and BLAS. The benchmark implementation is pinned to the same revision.

Setting	Recorded value
Threads	6
Batch / micro-batch	512 / 128
Requested GPU layers	99
Flash attention	Enabled
Key / value cache types	F16 / F16
KV offload	Enabled
Prior context	0 tokens
Requested context	512 tokens for pp512; 128 for tg128
Actual allocated context	Not retained
Warmup	One full prompt-processing pass before pp512; one single-token decode step before tg128
Timed work	Synthetic decode work and backend synchronization; model loading excluded
Repetitions	Three sequential repetitions per model/workload

For each repetition, throughput equals token count × 1,000,000,000 ÷ elapsed nanoseconds. The mean is taken across the resulting rates, rather than dividing token count by mean duration. Sample standard deviation uses a denominator of two for the three repetitions. All 18 original nanosecond measurements are included.

Exact model artifacts

Model	Quantization	Immutable file reference	Bytes
MiniCPM5-2B	Q4_K_M	MiniCPM5-2B-Q4_K_M.gguf	1,561,318,368
Spark-X2.5-1.7B	Q4_K_M	Spark-X2.5-1.7B-Q4_K_M.gguf	1,107,457,856
Spark-X2.5-4B	Q4_K_M	Spark-X2.5-4B-Q4_K_M.gguf	2,600,224,352

The JSON and manifest include exact repository revisions and SHA-256 digests. Model files are linked, not redistributed. Model artifacts retain their own upstream licenses, separate from any license for these measurements.

Limits

• These are historical llama.cpp engine microbenchmarks on one recorded hardware configuration. They do not measure Loci app performance or customer usage.
• The host was under substantial concurrent load: the recorded one-minute load averages before the three commands were approximately 31 to 33 on a 10-core CPU. Background work, thermal state and power mode were not controlled in the retained evidence.
• Each test has three sequential repetitions from one session on 30 September 2026. Variation across those repetitions does not describe variation across machines, people or future runs.
• The benchmark feeds synthetic token IDs. The generation test performs 128 single-token decode steps without a preceding prompt; it is not a 128-token natural-language answer or a quality test.
• Requested context was 512 tokens for prompt processing and 128 tokens for generation. Actual allocated context, OS version, exact Mac product identifier and peak process memory were not retained.
• The timings exclude model loading and include no app interface, time-to-first-token, tools, retrieval or conversation history. No phone or tablet was measured.
• Original timing and checksum receipts are retained. The original executable and model cache are no longer available for a fresh local hash check. The runtime commit and immutable model artifact hashes and sizes were independently checked against the upstream sources on 30 October 2026.
• No memory-fit threshold, Loci compatibility claim, device-wide speed estimate or cross-model quality conclusion follows from these measurements.

Files

• measurements.csv: One row per timed repetition, including token counts, elapsed time, recalculated throughput, exact model/runtime identity and known test conditions. Empty memory and first-token fields mean unrecorded, not zero.
• data.json: Original timing arrays, summaries, machine, protocol, model artifacts, retained receipt fingerprints and limitations.
• manifest.json: Edition identity, public source links, checksums, derived summaries and citation.
• SHA256SUMS: SHA-256 checksums of the distributed files.

No customer conversations, prompts, documents, installation identifiers or user records are included.

Citation

Loci AI, Inc. (2026). M1 Pro engine microbenchmarks, 30 September 2026. Three Q4_K_M models; synthetic pp512 and tg128 tests; three repetitions per test on one machine. https://askloci.ai/speed/lab/2026-09-30/

When quoting a result, include the model variant, workload, hardware, measurement date and three-repetition sample size. Identify the measurement as a historical synthetic engine run.

Reuse

No dataset reuse license is specified. Model licenses apply separately to the model artifacts.