I trained a 125M model to autocomplete piano on-device
Details
- External ID
- 49373456
- Source
- HN
- Company
- —
- Product
- I trained a 125M model to autocomplete piano on-device
- Website domain
- simedw.com
- Launched
- Aug. 20, 2026
- Cohort
- —
- Upvotes
- 598
- Upvotes percentile
- 0.9973118279569892
- Tags
- —
- Fetched at
- Sept. 10, 2026, 5:32 a.m.
- Updated at
- Sept. 10, 2026, 5:32 a.m.
Description
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15).The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device.The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
Enrichment
- Theme
- music creation and learning tools
- Vertical
- Media & entertainment
- Function
- Model & infra
- Audience
- Developer
- AI stance
- AI-native
- Project type
- Hobby / open-source project
- Normalized one-liner
- on-device piano autocomplete model
- Manually corrected
- False
Could you build this?
No Training a custom 125M-parameter autoregressive transformer on tokenized MIDI sequences, fine-tuning it with DPO, and optimizing it for low-latency on-device Apple Neural Engine (CoreML) inference requires specialized ML research and systems engineering.
What it would actually take: The architecture involves a custom byte/event-level MIDI tokenization scheme (like Remi or MIDI-BERT), training a custom 125M-parameter Transformer model using PyTorch on GPU clusters, and applying Direct Preference Optimization (DPO) to align musical output. For on-device real-time performance (~108 notes/sec), the model must be quantized and converted to Core ML / Metal Performance Shaders with custom low-latency KV-caching. This requires specialized deep learning, music information retrieval (MIR), and mobile hardware optimization expertise.
Discussion
20 comments analyzed.
Competitors mentioned: Ludwig (rule-based system by Fritz chess engine maker), music21, tunesage.com, CRIM, VIS, and humdrum (musical syntax analysis tools)
Concerns raised: Model produces non-idiomatic voice-leading and cadence patterns compared to classical era conventions, Creative collapse tendency - LLMs prefer averages rather than exploration, May struggle with proper phrase structure and harmonic development, Scalability concerns with complex instrumental techniques (bends, vibratos, harmonics) expanding event vocabulary
Feature requests: Web MIDI integration to run in browser, VST or Max 4 Live device implementation, Independent representation of different musical voices/parts, Modeled tie representation to reduce durational vocabulary, Fine-tuning capability on user's own MIDI data
Competitors
Other products that read as similar to this one — 116 launches clear the similarity bar, closest 8 shown.
Attention rank: #2 of 117 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 295 days after the earliest competitor.
- Public Muscriptor Instance (latest, most powerful Audio-to-MIDI model) · hn · 2026-08-21 · 58 upvotes · similarity 0.49
- I made a memory game to teach you to play piano by ear · hn · 2026-01-09 · 565 upvotes · similarity 0.47
- Clef · hn · 2026-08-07 · 8 upvotes · similarity 0.46
- lumikey · ph · 2026-09-17 · 15 upvotes · similarity 0.45
- LumaKeys · ph · 2026-09-23 · 7 upvotes · similarity 0.44
- Musical Interval Trainer · hn · 2026-02-11 · 22 upvotes · similarity 0.43
- MotifPilot MIDI Generator · ph · 2026-09-10 · 2 upvotes · similarity 0.43
- jevthoven · github · 2026-09-17 · 10 upvotes · similarity 0.42
Other launches for this product
- No other launches for this product.
Same idea, different domain
Nobody's really built a model & infra tool for Fintech yet.