Exploring Apple's On-Device Model: Twenty Billion Parameters in Flash
Comments on “Exploring Apple's On-Device Model: Twenty Billion Parameters in Flash”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Apple's Core Advanced is the larger of its two on-device models, with 20 billion parameters of which only 1 to 4 billion are active at once, built with Google and powering Siri's new voices and dictation.
Why it ranked
No. 35 of 62 with a score of 0.60. The panel rated it highest on technical depth (0.95) and new information (0.93), lowest on audience breadth (0.57). The claim check took 0.100 off.
Every answer behind this is in the table under Score.
Score
0.60total under rubric v2.2
- Significance0.66
- Reader interest0.83
- Audience breadth0.57
- Practical relevance0.70
- Substantiated0.88
- Technical depth0.95
- New information0.93
- Counts against: Promotional0.04
- Counts against: Overclaim0.29
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.62 | 0.55 | 0.82 | 0.66 | 0.27 | +0.24 | +0.159 |
| Reader interest | 0.74 | 0.80 | 0.95 | 0.83 | 0.21 | +0.26 | +0.216 |
| Audience breadth | 0.49 | 0.50 | 0.72 | 0.57 | 0.23 | +0.13 | +0.074 |
| Practical relevance | 0.68 | 0.60 | 0.83 | 0.70 | 0.23 | +0.22 | +0.155 |
| Substantiated | 0.94 | 0.80 | 0.90 | 0.88 | 0.14 | +0.05 | +0.044 |
| Technical depth | 0.99 | 0.90 | 0.95 | 0.95 | 0.09 | +0.05 | +0.047 |
| New information | 0.91 | 0.90 | 0.99 | 0.93 | 0.09 | +0.05 | +0.047 |
| Promotional | 0.07 | 0.05 | 0.01 | 0.04 | 0.06 | −0.19 | −0.008 |
| Overclaim | 0.37 | 0.30 | 0.20 | 0.29 | 0.17 | −0.11 | −0.032 |
| Sum of contributions | +0.702 | ||||||
| Claim-check penalty | −0.100 | ||||||
| Total, as published | 0.602 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-10-26/35/"><img src="https://rubric.hrih.si/badge/2026-10-26/35.svg" alt="Ranked #35 on Rubric, 26 Oct 2026" width="245" height="28"></a>Claim check
Claim check: 5 matches source · 1 differs from source · 1 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
Apple's Core Advanced on-device model has 20 billion parameters, of which 1 to 4 billion are active at a time, and Apple says it is custom-built in collaboration with Google.
The cited source says the same thing.
The article says “with 20 billion parameters of which 1 to 4 billion work at a time. Apple says it is "custom-built in collaboration with Google”
The source says “this 20-billion-parameter model uses a sparse architecture, activating just 1 to 4 billion parameters at a time depending on the request”
Source cited by the article: machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models
Differs from sourceNumber
Two seconds after a long request, Apple's model and its safety models held 2.0 GiB of memory, while an open model of the same class built the usual way held about 15 GiB.
The cited source says something else; both quotes are shown.
The article says “Apple's model and its safety models held 2.0 GiB of memory; an open model of the same class, built the usual way, held about 15 GiB”
The source says “2.0 GiB with the second process that runs the safety models (140 plus 535 MiB) | the open builds' `mlx_lm.server` processes: 3.4 to 4.6 GiB at 4 bits, 5.6 to…”
Source cited by the article: github.com/juliensimon/apple-fm-mac/tree/main/findings/04-memory
Couldn’t checkNumber
On the same 20 math problems, Apple's model used about a third of the energy of the best open build the author ran.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “On the same 20 math problems, it used about a third of the energy of the best open build I ran”
The source says No passage could be compared.
Source cited by the article: github.com/juliensimon/apple-fm-mac/tree/main/findings/05-power-and-energy
No source givenBenchmark result
On 100 grade-school math problems Apple's model solved 78 while the three open builds solved 89 to 94, a gap too large for chance.
The article gives no link or citation for it.
The article says “On 100 grade-school problems, Apple's model solved 78 and the three open builds 89 to 94, a gap too large for chance”
The source says No source to compare with.
Matches sourceNumber
Google's Gemma 4 26B A4B stores 25.2 billion parameters and runs 3.8 billion per token, using 8 of its 128 experts plus a shared one.
The cited source says the same thing.
The article says “Google's Gemma 4 26B A4B stores 25.2 billion parameters and runs 3.8 billion for each token, 8 of its 128 experts plus a shared one”
The source says “إجمالي المَعلمات | 25.2 مليار | | المعلَمات النشطة | 3.8 مليار”
Source cited by the article: ai.google.dev/gemma/docs/core/model_card_4
Matches sourceNumber
Apple's fm count-tokens matched Gemma 3's token count on all 429 strings tried, while tokenizers from other labs matched on less than half.
The cited source says the same thing.
The article says “Apple's fm count-tokens gave Gemma 3's count on all 429 strings I tried, from code to sentences in 18 languages; tokenizers from other labs matched on less than half”
The source says “Gemma 3 429, Gemma 3n 429, Gemma 4 425, Gemma 2 324 (75.5%); eight tokenizers from other labs 20.7% to 44.5%”
Source cited by the article: github.com/juliensimon/apple-fm-mac
Matches sourceQuote
Apple says Core Advanced makes routing decisions per prompt because NAND-to-DRAM bandwidth is too slow to swap weights token by token as standard MoE models require.
The cited source says the same thing.
The article says “NAND-to-DRAM bandwidth is too slow to swap weights token by token, as standard MoE models require." So Core Advanced "makes routing decisions per prompt”
The source says “Because NAND-to-DRAM bandwidth is too slow to swap weights token by token, as standard MoE models require, AFM 3 Core Advanced makes routing decisions per prompt”
Source cited by the article: machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models
Matches sourceQuote
An Apple engineer wrote that there is currently no pinning or model version retrieval API for the on-device model.
The cited source says the same thing.
The article says “an Apple engineer wrote that there is "currently no pinning or model version retrieval API”
The source says “There is currently no pinning or model version retrieval API”
Source cited by the article: developer.apple.com/forums/thread/833642
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 26 Oct 2026, 07:27 UTC. It played no part in the ranking.