Qwen2.5-1.5B Instruct / 28L LUT4, production ANEMLL Core ML conversion
MLComputePlan / infer2,090 ANE / 79 CPU
CPU residual: lm_head / argmax tail. 96.4% / 3.6% by assigned ops. Evidence
Public research summaryTHE PUBLIC RECORD / 01
MacBrains Labs / Evidence explorer
Find the claim.
Follow the evidence.
Search the measurements, field notes, and verified archive releases already published by MacBrains Labs. Every entry links back to its public source.
PUBLIC EVIDENCE INDEX / 02
What is actually on record.
Measurement, compiler assignment, interpretation, and archive integrity are different kinds of evidence.
13 entries from published sources
Qwen2.5-1.5B / KV state path
ANE-assigned state primitives394 / 394
56 slice_update + 57 read_state + 56 write_state + 225 slice_by_index.
Public research summaryQwen2.5 / decode batch sweep
Aggregate tokens per second102.1 at B=4
Five-round interleaved steady-state run on a single-function, batch-shaped probe. Throughput only; output quality not validated.
Public research summaryQwen2.5 / LUT4 model quality
WikiText-2 perplexity16.8033
Hand-authored 28L quantized model, run as two 14L halves sharing KV, with position-correct RoPE. Full 12,000-token validation set.
Public research summaryLFM S2 sealed production graph
ANE compute placement743 / 743
MLComputePlan for two-output artifact, not separately compiled debug graph.
Public research summaryLFM S2 / output parity
ANE vs fixture relative L20.01730 / 0.02786
Two declared production outputs. ANE gate relative L2 ≤ 0.03.
Public research summary28 layers, LUT4, and the cost of exact placement
Research field noteQwen2.5 · MIL · KV state
State reads, indexed slices, weight.bin offsets and the non-ANE tail.
Public field note N/001A passing production graph and a failing diagnostic one
Research field noteLFM · S2 · Parity
LFM block-pair parity, 743/743 assignment and why probes alter the experiment.
Public field note N/002The bandwidth limit that was not a limit
Research field noteA18 Pro · Decode
Batch sweeps, dispatch floors and a hypothesis overturned by measurement.
Public field note N/003Every slice passed. The chain did not.
Research field noteLFM2.5 · Parity · Negative
LFM2.5-1.2B layer slices meet a strict ANE parity gate; the 16-layer chain and token logits miss it.
Public field note N/004Correct, placed, paid for
Research field notePlacement · Dispatch · Grafts
A ring-buffer attention graft placed nothing on the ANE; an EAGLE draft head made every call slower and has to earn it back.
Public field note N/005LFM3-1.7B layer-14 diagnostic under Core ML cpuAndNeuralEngine (same chip, single device)
Published archive integrity: verifiedv0.2.0
same-chip, byte-reproducible on Apple A18 Pro (Mac17,5), macOS 27.2 (build 26B123), CoreML 3600.25.2; ANE execution INFERRED from Core ML compute-plan placement; no novelty claim
Published release and DOI 10.5281/zenodo.23905811MacBrains Labs ANE Research Records, 2026: Public Measurement Summaries
Published archive integrity: verifiedv0.1.0
source-bound curated summary, not raw results; archive integrity is not independent replication or peer review
Published release and DOI 10.5281/zenodo.23798037No published records match these filters. Try a broader search.
HOW TO READ IT / 03
Evidence isn't a single badge.
Planner reported: Core ML's assigned compute unit, not a physical trace of the Neural Engine.
Measured: an observed result for a named model, device, compiler and method. It does not generalize automatically.
Field note: engineering interpretation, including counterexamples and failed hypotheses.
Verified archive: release provenance and file integrity. A DOI does not imply peer review or independent replication.
Read the full research record ↗