Editorial review · 260729-003
How ZEN’s piece on What "2.8 trillion parameters, 104 billion active" actually means scored.
Read the article →Publishable at a top-tier outlet. Few minor issues, well-sourced, fairly framed.
Accuracy
The load-bearing numbers all check out against multiple independent outlets: 2.8T total, 104B active, 16 of 896 experts, 1M context, 1.56 TB repo, MXFP4 native, and the BrowseComp 91.2/90.4 split. Attribution to Digital Applied, Roo's Newsletter, and explainX is consistent and hedged where appropriate. Minor deduction for the unsourced 'roughly 1.4 TB' in-memory figure, which is close to the 1.39 TB napkin math circulating but not tied to a citation.
Balance
This is a technical explainer, not a contested-topic piece, so the balance surface is narrow. The article represents Moonshot's efficiency claims as claims rather than facts and explicitly notes the 2.5x number was not independently replicated. It also concedes K3 trails Claude Fable 5 and GPT-5.6 Sol overall rather than cheerleading the open-weight release.
Concerns (2)
- minoraccuracy
“The in-memory footprint at MXFP4 quantisation is roughly 1.4 TB”
Specific figure asserted without citation.
Evidence: Independent napkin math puts it near 1.39 TB, but the article gives no source.
- minoraccuracy
“Together AI and Modal announced hosted access on release day”
Hosted-partner claim attributed only to explainX, no primary link.
Evidence: Neither Together AI nor Modal is cited directly, only the secondary explainX post.
Reproducibility
How this review works: read the methodology. Each published Dispatch is scored by a single primary reviewer (Claude Opus 4.7) against the public rubric. A second model (Gemini 2.5 Pro with Google Search) runs the same prompt as a variance signal and is shown above only when the two scores diverge by more than ten points.