Nicheloom

Market intelligence for builders — see what's gaining traction before it's crowded.

LoRA gradients on Apple's Neural Engine at 2.8W

Details

External ID
47277219
Source
HN
Company
—
Product
LoRA gradients on Apple's Neural Engine at 2.8W
Website domain
github.com
Launched
March 6, 2026
Cohort
—
Upvotes
6
Upvotes percentile
0.2853628536285363
Tags
—
Fetched at
Sept. 7, 2026, 9:26 p.m.
Updated at
Sept. 7, 2026, 9:26 p.m.

Enrichment

Theme
ML inference and model optimization
Vertical
Horizontal
Function
Model & infra
Audience
Developer
AI stance
AI-native
Project type
Hobby / open-source project
Normalized one-liner
lora training on apple neural engine
Manually corrected
False

Could you build this?

No Writing backpropagation and gradient calculation routines to run on Apple's proprietary, undocumented Neural Engine (ANE) requires reverse-engineering hardware assembly/IR and specialized low-level ML compiler skills.

What it would actually take: Requires reverse-engineering Apple's proprietary MIL (Model Intermediate Language) compiler and low-level ANE instructions via PrivateFrameworks. The implementer must formulate custom backward-pass kernels for matrix multiplication and LoRA updates that bypass CoreML's forward-only constraints, requiring expert compiler and hardware reverse-engineering knowledge.

Discussion

1 comment analyzed.

Competitors mentioned: MLX, maderix's ANE reverse engineering

Concerns raised: ANE compiler leaks handles, silently fails after ~119 compiles, ANE matmul op compiles but never executes, Spatial dimensions must be multiples of 16

Competitors

Other products that read as similar to this one — 1628 launches clear the similarity bar, closest 8 shown.

Attention rank: #1069 of 1629 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).

Launched 128 days after the earliest competitor.

Other launches for this product

Same idea, different domain

Nobody's really built a model & infra tool for Fintech yet.