Loci speed data / 2026-10-30
The measurement, the sample and the method belong together. Explore how these observations are collected and what a fair comparison requires.
Inspect the raw repeats and exact model files from a historical llama.cpp test on one M1 Pro Mac. Engine measurements are separate from Loci app speed.
Fixed edition · 2026-10-30
Observation window: to . Methods 1.0.0.
The rolling speed page changes as observations arrive. A fixed edition keeps its source, dates and methods together.
How to read the numbers
The source records timings during ordinary Loci conversations. Prompts, answer lengths, conversation history and device conditions vary. These observations describe the recorded sample; they do not isolate hardware performance or model quality.
The exporter first summarizes observations within each installation, then takes the median across installations for a device-and-model group. Phones summarize completion and continuation events; Macs summarize run medians. One busy installation does not outweigh many quieter ones. Installation counts are not counts of people.
Phone throughput uses four output characters as one estimated token. Mac throughput counts the model’s tokens. These units are identified separately in the research export. They must not be combined into a phone-versus-Mac speed ranking.
The 25th and 75th percentiles describe the middle half of installation-level medians. They are not a confidence interval or a promise about another device. Research exports require at least ten installations in a device-and-model group.
The reported delay before generation starts is distinct from the speed of the answer once it starts. Longer prompts, model loading and runtime behavior can affect the delay. A median delay plus a median throughput does not describe a particular completed answer.
A citable performance edition requires source checks for public-release eligibility, internal and test exclusions, delivery coverage and the full observation window. A successful export alone does not establish those properties.
The current aggregate does not identify an exact model artifact, quantization, runtime revision, context length or peak memory for every row. Model names alone do not establish equivalent settings. The data cannot determine whether an untested model fits in memory.
Phone filters exclude reported low-power, background and tool-assisted activity, but some missing state fields pass the original filters. That does not establish that every observation used the same settings. Source counts also differ: completion and continuation events on phones, reported answer counts within runs on Mac.
Research exports contain aggregate measurements. Conversation text and installation identifiers are excluded.
| Field | Meaning |
|---|---|
edition | Fixed edition identifier; retained when the rolling speed page changes. |
period_start / period_end | Observation-window labels reported by the source exporter, in UTC. |
device / chip / platform | Named device configuration and platform; chip variants remain separate. |
model_id / model_name | Recorded model identity. Exact artifact and runtime may be unknown. |
speed_unit | character_estimated_tokens_per_second on phones; model_tokens_per_second on Mac. |
speed_median / speed_p25 / speed_p75 | Median and middle-half spread across qualifying installation summaries. |
first_token_ms | Median recorded first-token delay in milliseconds, or empty when unavailable. |
installations / observations / observation_unit | Installation counts and source-specific observation counts; never summed into unique people or treated as unique saved answers. |
aggregation | The installation-level summary used before the across-installation median. |
artifact_revision / quantization / runtime / context_tokens | Empty where the original aggregate cannot establish the value. |
A stable reference
Loci AI, Inc. (2026). Local AI speed observations, edition 2026-10-30. Methods 1.0.0. https://askloci.ai/speed/research/2026-10-30/
Cite this edition for its methods and documented limitations. It is not a source for verified performance comparisons.
3eb9c70c4aef6fa298dd6b9f76383bad886b499a771b513e3874502c0f64fc2c7c3368b920d0b5876878c2402d192ba6d9fec510Check model requirements and see where the evidence is strong enough to guide a choice.