Apple's AFM 3 stores 20B weights in flash, with deployment details pending
Apple's AFM 3 Core Advanced demonstrates that storing a 20-billion-parameter language model in NAND flash rather than DRAM is technically tractable, but the production case rests on metrics the company has not yet published. The architecture lifts the on-device capability ceiling but leaves the deployment questions that gate enterprise procurement outside the announcement.
According to Apple, AFM 3 Core Advanced stores its full 20-billion-parameter weight set in NAND flash and uses a routing scheme the company calls Instruction-Following Pruning, developed with Apple's own researchers. Because NAND-to-DRAM bandwidth is too slow to swap experts token by token as standard Mixture-of-Experts models require, the model routes once per prompt, selects a fixed expert set, and generates all tokens from that configuration. Active parameters scale from 1 billion for simpler operations, up to 4 billion for harder ones, drawn from the 20-billion-parameter pool in flash alongside always-active shared experts that remain in DRAM. The total pool is the storage footprint; the active count is what consumes energy and bandwidth during inference.
Hannun's X posts describe the routing from an applied perspective. Awni Hannun, a researcher at Anthropic and a former Apple research scientist, posted that "you can't put 20B parameters in RAM at any reasonable precision," and described Apple's approach as "pretty exotic architecture by today's standards." He added that "a small model predicts from the query (or prompt) which experts to load from NAND into RAM," and that "the key distinction from a typical MoE is that you do this once per query and then generate all the tokens with the same experts." Hannun's framing places the routing decision before generation rather than during it, the only ordering that allows NAND to substitute for DRAM at this scale.
On-device deployments have been parameter-capped by the size of active memory, which is why practical on-device LLMs have stayed in the low billions. The architecture shifts that constraint: Apple's 20-billion-parameter pool fits within a budget envelope that includes NAND, not just DRAM. The design claim is real, and per-prompt routing is the only way it can work given NAND's bandwidth ceiling relative to DRAM. Production viability now depends on device thermal design and memory subsystem behavior rather than on DRAM capacity alone.
The design description does not address what on-device deployment demands at runtime. Marco Abis, who is building Ziraph, a profiler for local AI on Apple silicon, posted on X that Apple's documentation does not include "energy, memory bandwidth, thermal" data and called out the absence as a "notable gap, given those decide most of on-device performance." Apple's profiling tools expose timing, but the source does not specify what the architecture costs in energy, thermal headroom, or memory bandwidth under sustained load, which is the input needed to size devices, write thermal policies, and reason about throttling and battery impact under real user behavior.
Routing between on-device and server-side models is undocumented in the materials Abis reviewed. Abis noted that Apple's Core AI docs, Foundation Models docs, and Private Cloud Compute security post do not state when an on-device request transparently offloads to AFM 3 Cloud Pro, or whether that routing decision is visible to developers or end users. Documentation requirements for inference location, common in regulated industries, cannot be satisfied without a developer-facing signal. The architectural choice between on-device and Private Cloud Compute is now a runtime decision the calling code cannot directly observe, and any audit log of inference location is necessarily incomplete without one.
Server-side inference is where the agentic workload runs at scale. According to the source, AFM 3 Cloud Pro runs on Nvidia GPUs in Google Cloud and handles complex agentic tool use and reasoning. The data-privacy guarantee travels with Private Cloud Compute, but the Google Cloud dependency does not. Enterprises that prefer not to run agentic inference on hyperscaler infrastructure, or that have data-residency constraints that complicate Google Cloud routing, inherit that dependency along with the privacy envelope. The 20-billion-parameter on-device option does not displace the server-side commitment; it sits beside it.
Apple has indicated that a full technical report with benchmarks is due later this summer. That report is the natural place for the missing numbers. Whether the report closes the energy, thermal, and routing-transparency gaps or confirms them, the procurement question has moved past model size to operating limits Apple has not yet disclosed.