Independent testing will determine whether Maple-Preview’s ternary 20B MoE sustains roughly 120 tokens per second on an iPhone while retaining practically useful model quality.

What is this?

Why it matters to Scott

Measured heat

How the heat travelled

Evidence (5) — ⭐ canonical anchor

Interpretation history

• 2026-08-20T19:39:31Z
  No substantive validation followed the underspecified iPhone report or pending runtime-support route, and the speed-plus-quality claim has faded beyond its active evidence window. Expire this episode; merged support or reproducible independent iPhone testing would constitute a fresh development.
• 2026-08-18T18:58:47Z
  The staleness trigger adds no evidence; the case remains supported only by a third-party desktop implementation and an underspecified iPhone anecdote, while practical quality is still doubtful. Keep an infrequent watch for merged runtime support and reproducible speed-plus-quality testing.
• 2026-08-16T18:35:41Z
  Refreshed comments add skepticism about benchmark comparisons and intended tool-routing use, but no reproducible Maple-specific iPhone measurement, merged runtime support, or representative quality test. The case remains a cold validation watch pending upstream support and independent speed-plus-quality benchmarking.
• 2026-08-16T16:34:24Z
  Refreshed comments add no Maple-specific measurement, merged runtime support, device configuration, or representative quality test; discussion of another ternary model does not validate this claim. The case remains a cold watch for reproducible iPhone benchmarking after llama.cpp support lands.
• 2026-08-16T13:28:14Z
  The refreshed discussion points to a pending llama.cpp support PR, providing a concrete route toward reproducibility but no merged implementation, device configuration, new measurement, or practical-quality evaluation. It does not materially advance the earlier anecdotal iPhone result.
• 2026-08-16T10:30:46Z
  A first-person report of roughly 100 tokens/s makes the headline iPhone throughput more plausible alongside the M4 Air implementation, but it lacks device details, configuration, reproducibility, and representative quality testing. It is useful directional evidence, not enough substance to corroborate the combined speed-and-quality hypothesis.
• 2026-08-16T10:23:11Z
  evidence attached: reddit.post.1vpsozs — The reported Maple ternary model reaching roughly 100 tokens per second on an iPhone is independent user evidence for the open edge-inference case.
• 2026-08-14T15:46:48Z
  No substantive evidence has appeared since the M4 Air implementation; the headline iPhone throughput and practical quality remain unvalidated. The case is now an infrequent benchmark watch rather than an active developing signal.
• 2026-08-12T14:53:46Z
  The refreshed discussion and minor engagement add no reproducible iPhone benchmark or representative quality evaluation. The M4 Air implementation remains the sole independent implementation result, so this should stay a cold, infrequent validation watch.
• 2026-08-10T13:34:05Z
  Refreshed discussion adds a plausible latency-bound explanation for the Mference result but no new measurements, independent iPhone test, or practical-quality evaluation. The case remains a cold validation watch rather than advancing beyond the existing third-party implementation.
• 2026-08-10T10:25:55Z
  The Mference integration is the first independent implementation evidence that Maple-Preview can run with unusually low reported memory use, reaching about 40 tokens/s on an M4 Air. It does not validate the claimed 120 tokens/s on an iPhone or resolve practical-quality concerns, so the core hypothesis remains open.
• 2026-08-10T10:22:08Z
  evidence attached: reddit.post.1vkglog — The reported Mference integration and 40-token-per-second result add practical local-inference evidence for Maple-Preview.
• 2026-08-10T07:29:08Z
  No new evidence beyond marginal engagement drift; the same hallucination anecdotes remain the only counterevidence and no independent iPhone throughput or reproducible quality benchmark has appeared. Case stays a cold, unresolved validation watch.
• 2026-08-08T07:22:30Z
  Fresh hands-on comments independently report confident factual errors and aggressive hallucination, adding early counterevidence on practical quality even though the claimed iPhone throughput remains unverified. These anecdotes justify a watching state, but not corroboration without reproducible speed and quality benchmarks.
• 2026-08-06T06:24:11Z
  The attachment adds no identifiable independent benchmark or practical-quality evaluation, so the case remains an unverified first-party claim despite continued attention. Engagement-only triggers are exhausted; revisit only on reproducible third-party iPhone testing.
• 2026-08-06T05:24:05Z
  The attachment trigger exposes no identifiable new artifact or independent testing, so the case remains an unverified first-party performance claim. Repetitive engagement updates are exhausted; revisit only when reproducible iPhone throughput and practical-quality results appear.
• 2026-08-06T04:27:49Z
  No identifiable independent benchmark or quality evaluation was added; this is another engagement-driven reobservation of the same first-party claim. The case remains a cold validation watch, with further review warranted only if reproducible iPhone conditions, throughput, and practical-quality results emerge.
• 2026-08-06T03:26:19Z
  The trigger exposes no identifiable new artifact or independent testing; it is another engagement-driven reobservation of the same first-party claim. Keep only a cold, infrequent watch for reproducible iPhone throughput, operating conditions, and practical-quality results.
• 2026-08-06T01:25:31Z
  The trigger exposes no new identifiable evidence; the attached material still lacks an independent, reproducible iPhone benchmark and practical-quality evaluation. Repetitive amplification is exhausted, so retain only a cold validation watch.
• 2026-08-05T22:24:30Z
  The trigger exposes no identifiable independent benchmark or quality evaluation beyond the existing first-party claim and amplification. The case remains a cold validation watch; engagement-only reobservations no longer justify frequent review.
• 2026-08-05T19:33:55Z
  The trigger exposes no identifiable new evidence or independent testing, so the case remains an unverified first-party performance claim. Repeated engagement updates are informationally exhausted; wait for reproducible iPhone benchmarks and practical-quality evaluation.
• 2026-08-05T16:32:22Z
  The trigger exposes no identifiable independent testing beyond the already-known claim and amplification, so the case’s meaning remains unchanged. Repeated engagement-only updates are exhausted; revisit only when reproducible iPhone throughput, operating conditions, and practical-quality results emerge.
• 2026-08-05T14:28:42Z
  The attachment adds no identifiable independent benchmark or quality assessment beyond the existing claim and amplification. Despite a hot local-inference topic, this case remains a cold validation watch pending reproducible iPhone throughput, operating conditions, and practical-quality results.
• 2026-08-05T13:29:53Z
  The trigger supplies no identifiable independent benchmark or quality evaluation, so it does not change the case beyond the original first-party claim. Repetitive amplification is exhausted; retain a cold watch for reproducible iPhone throughput, operating conditions, and practical-quality testing.
• 2026-08-05T12:23:48Z
  The new trigger contains no identifiable independent benchmark or quality evaluation and only reobserves existing amplification. Keep the case cold until reproducible iPhone conditions, throughput, and practical-quality results appear.
• 2026-08-05T11:27:40Z
  The trigger adds no identifiable independent test; discussion remains repetitive amplification of the original claim, with engagement now flat. Keep the case cold and widen the review interval until reproducible iPhone speed, operating conditions, and practical quality results appear.
• 2026-08-05T10:23:18Z
  The latest trigger adds no identifiable independent testing and only reobserves the existing claim. Repetitive amplification has exhausted its value; keep this as a cold benchmark watch until reproducible iPhone speed, operating conditions, and practical quality results emerge.
• 2026-08-05T09:25:50Z
  The latest trigger adds no substantive evidence beyond the already-known first-party claim and its amplification. Without a reproducible independent iPhone benchmark or practical quality evaluation, the case remains a cold validation watch.
• 2026-08-05T08:28:00Z
  The trigger contains no identifiable new evidence beyond repeated observation of the existing posts, so it does not advance independent validation. The case remains a cold benchmark watch until reproducible iPhone performance and practical quality results appear.
• 2026-08-05T07:22:46Z
  The latest trigger adds no identifiable independent benchmark, reproducible iPhone conditions, or practical quality evaluation. Repeated amplification has exhausted its informational value, so the case should remain cold pending substantive third-party testing.
• 2026-08-05T06:27:15Z
  The new attachment provides no independent benchmark, reproducible device conditions, or practical quality assessment. Activity remains repetitive amplification, so the case should stay cold until substantive third-party testing appears.
• 2026-08-05T05:24:12Z
  The newly attached material still does not provide an independent iPhone benchmark, reproducible operating conditions, or practical quality testing. The case remains another unverified instance of a familiar local-inference pattern, so further amplification does not increase its significance.
• 2026-08-05T04:23:15Z
  The purported new evidence adds no independent benchmark, device conditions, or practical quality assessment; rising engagement remains repetitive amplification. The claim stays testable, but the case should wait for substantive third-party validation.
• 2026-08-05T03:28:13Z
  The latest attachment still provides no independent device benchmark, operating conditions, or practical quality evaluation. This is repetitive amplification of the original claim, so the case remains a low-priority validation watch.
• 2026-08-05T02:29:42Z
  No independent benchmark or quality evaluation has appeared; the new activity only repeats the first-party artifact and claim. The case remains testable but has not advanced, so attention can cool pending substantive validation.
• 2026-08-05T01:22:13Z
  The latest activity remains amplification of the same first-party artifact rather than an independent device benchmark or quality evaluation. The unusually concrete claim stays testable and worth watching, but its meaning has not advanced beyond an unverified local-inference result.
• 2026-08-05T00:26:16Z
  The attached model announcement identifies the underlying artifact but remains amplification of the first-party claim, not independent validation. Speed, device conditions, and practical quality therefore remain unsettled despite modest attention growth.
• 2026-08-05T00:21:27Z
  evidence attached: reddit.post.1vfrr5t — The announced 20B-A1B ternary model is the underlying artifact for the open case on useful high-speed iPhone inference.
• 2026-08-04T23:24:38Z
  grounded: known/low — The case adds no new position beyond Scott’s Capability Audit and hardware-aware local inference work: claimed mobile throughput still needs representative, ind
• 2026-08-04T23:22:03Z
  case created — The first-party page makes an unusually strong, concrete edge-inference claim that can be independently benchmarked for speed and quality.

Decision trace

• 08-21 05:39expireNo substantive validation followed the underspecified iPhone report or pending runtime-support route, and the speed-plus-quality claim has faded beyond its active evidence window. Expire this episode;
• 08-21 05:39alert_silentThe delta is only a staleness trigger with no new event, benchmark, implementation, or quality evidence, so there is nothing consequential to surface.
• 08-21 05:39alert_routeThe delta is only a staleness trigger with no new event, benchmark, implementation, or quality evidence, so there is nothing consequential to surface.
• 08-19 04:58repriceThe staleness trigger adds no evidence; the case remains supported only by a third-party desktop implementation and an underspecified iPhone anecdote, while practical quality is still doubtful. Keep a
• 08-19 04:58alert_silentNo new event or consequential evidence occurred, so this can wait for merged llama.cpp support or a reproducible independent iPhone benchmark with representative quality testing.
• 08-19 04:58alert_routeNo new event or consequential evidence occurred, so this can wait for merged llama.cpp support or a reproducible independent iPhone benchmark with representative quality testing.
• 08-17 08:21sensor_dirtyengagement_update
• 08-17 04:35repriceRefreshed comments add skepticism about benchmark comparisons and intended tool-routing use, but no reproducible Maple-specific iPhone measurement, merged runtime support, or representative quality te
• 08-17 04:35alert_silentThe comment refresh does not establish a new event or materially strengthen either performance validation or quality counterevidence; it can wait for the llama.cpp support PR to merge or a reproducibl
• 08-17 04:35alert_routeThe comment refresh does not establish a new event or materially strengthen either performance validation or quality counterevidence; it can wait for the llama.cpp support PR to merge or a reproducibl
• 08-17 04:21sensor_dirtycomment_update
• 08-17 02:34repriceRefreshed comments add no Maple-specific measurement, merged runtime support, device configuration, or representative quality test; discussion of another ternary model does not validate this claim. Th
• 08-17 02:34alert_silentThe discussion only repeats the pending llama.cpp route and adds unrelated ternary-model experience; without a merged implementation or reproducible Maple benchmark, the delta can wait.
• 08-17 02:34alert_routeThe discussion only repeats the pending llama.cpp route and adds unrelated ternary-model experience; without a merged implementation or reproducible Maple benchmark, the delta can wait.
• 08-17 02:21sensor_dirtycomment_update
• 08-16 23:28repriceThe refreshed discussion points to a pending llama.cpp support PR, providing a concrete route toward reproducibility but no merged implementation, device configuration, new measurement, or practical-q
• 08-16 23:28alert_silentA comment linking pending upstream support is useful tracking context, but it neither establishes that support shipped nor adds reproducible speed or quality results; it can wait for the PR to merge o
• 08-16 23:28alert_routeA comment linking pending upstream support is useful tracking context, but it neither establishes that support shipped nor adds reproducible speed or quality results; it can wait for the PR to merge o
• 08-16 23:21sensor_dirtycomment_update
• 08-16 20:30repriceA first-person report of roughly 100 tokens/s makes the headline iPhone throughput more plausible alongside the M4 Air implementation, but it lacks device details, configuration, reproducibility, and
• 08-16 20:30alert_silentThe new report is an underspecified anecdote rather than a reproducible benchmark, so it does not change Scott’s local-inference decisions and can wait for routine briefing or stronger testing.
• 08-16 20:30alert_routeThe new report is an underspecified anecdote rather than a reproducible benchmark, so it does not change Scott’s local-inference decisions and can wait for routine briefing or stronger testing.
• 08-16 20:24alert_silentA low-engagement Reddit post adds only an unspecified first-person claim of roughly 100 tokens per second on an iPhone. It provides no device model, test configuration, quality measurements, or reprod
• 08-16 20:24alert_routeA low-engagement Reddit post adds only an unspecified first-person claim of roughly 100 tokens per second on an iPhone. It provides no device model, test configuration, quality measurements, or reprod
• 08-16 20:23attachThe reported Maple ternary model reaching roughly 100 tokens per second on an iPhone is independent user evidence for the open edge-inference case.
• 08-16 20:21propose_attachThe reported Maple ternary model reaching roughly 100 tokens per second on an iPhone is independent user evidence for the open edge-inference case.
• 08-15 01:46repriceNo substantive evidence has appeared since the M4 Air implementation; the headline iPhone throughput and practical quality remain unvalidated. The case is now an infrequent benchmark watch rather than
• 08-15 01:46alert_silentThis trigger is only a staleness check with negligible engagement drift and no new benchmark, implementation, or quality evaluation, so there is nothing consequential to surface before the next briefi
• 08-15 01:46alert_routeThis trigger is only a staleness check with negligible engagement drift and no new benchmark, implementation, or quality evaluation, so there is nothing consequential to surface before the next briefi
• 08-13 00:53repriceThe refreshed discussion and minor engagement add no reproducible iPhone benchmark or representative quality evaluation. The M4 Air implementation remains the sole independent implementation result, s
• 08-13 00:53alert_silentThe delta is repetitive discussion and marginal engagement rather than a new measurement, implementation, or release; it can wait unless an independent iPhone throughput and quality test appears.
• 08-13 00:53alert_routeThe delta is repetitive discussion and marginal engagement rather than a new measurement, implementation, or release; it can wait unless an independent iPhone throughput and quality test appears.
• 08-10 23:34repriceRefreshed discussion adds a plausible latency-bound explanation for the Mference result but no new measurements, independent iPhone test, or practical-quality evaluation. The case remains a cold valid
• 08-10 23:34alert_silentThe new comments are technical interpretation and use-case discussion, not a consequential new benchmark or release. They can wait for a reproducible iPhone throughput result or representative quality
• 08-10 23:34alert_routeThe new comments are technical interpretation and use-case discussion, not a consequential new benchmark or release. They can wait for a reproducible iPhone throughput result or representative quality
• 08-10 23:21sensor_dirtycomment_update
• 08-10 20:25repriceThe Mference integration is the first independent implementation evidence that Maple-Preview can run with unusually low reported memory use, reaching about 40 tokens/s on an M4 Air. It does not valida
• 08-10 20:25alert_silentThe third-party implementation is substantive but tests different hardware and leaves both headline iPhone throughput and representative quality unverified. It improves the case without changing Scott
• 08-10 20:25alert_routeThe third-party implementation is substantive but tests different hardware and leaves both headline iPhone throughput and representative quality unverified. It improves the case without changing Scott
• 08-10 20:22alert_silentA third-party Mference implementation reports a concrete Air M4 result of about 40 tokens/s with 500 MB RAM, but it neither tests the headline iPhone throughput nor provides representative quality eva