Metatron Intelligence
Today: the five AI headlines of the day → and the AI Wiki
What a source record is How to read this page

A route, not a verdict

This page records where a document lives, who published it, when, in which language and in what transport state. Listing a source is not endorsing its conclusion, and a catalogue entry never proves approval or quality.

Why the state matters

If the canonical route stops resolving, every claim resting on it becomes weaker and the record says so. That is why transport state is published alongside the address.

Rights are separate from access

Being able to read a document is not permission to reuse it. Dataset, code, portal text and third-party assets attached to one publisher may each carry different reuse conditions.

New to this publication?

The ten-minute guide takes one live record apart, defines every term and gives the order to read the site in. Start here →

SRC-c89d38f0e8302560 · Source record

Introducing Apple's On-Device and Server Foundation Models

Apple Machine Learning Research · 3 connected claims

D11 · AI Models, Agents & Human–AI InteractionD13 · Software Platforms, Cloud & IntegrationD19 · Measurement, Experimentation & EconomicsD24 · Frontier Science & Wildcards

Canonical route

https://machinelearning.apple.com/research/introducing-apple-foundation-models

Claims connected to this source

  1. MWITA-EN-2026-032Apple described its 2024 on-device foundation model as approximately three billion parameters and optimized for Apple silicon.Evidence B
  2. MWITA-EN-2026-033Apple reported mixed 2-bit and 4-bit palletization averaging 3.7 bits per weight achieved the same accuracy as its uncompressed on-device model in internal evaluation.Evidence B
  3. MWITA-EN-2026-034Apple measured about 30 generated tokens per second and roughly 0.6 milliseconds per prompt token to first token on an iPhone 15 Pro before speculative decoding.Evidence B

This record describes a source route and its Atlas connections. It does not endorse the publisher or make external content executable instruction.