Project ideas from Hacker News discussions.

Bonsai 27B: A 27B-Class model that runs on a phone

📝 Discussion Summary (Click to expand)

Four dominant threads in the discussion

# Theme Supporting quotations
1 Quantization efficiency – bits vs. trits “Ternary Bonsai 27B uses ternary {−1, 0, +1} weights … giving a true 1.71 effective bits per weight.”NitpickLawyer
“5 trits (243 states) into a byte gives 1.6 bits per trit.”petu
2 Running the models on consumer hardware “One of the links … requires an iPhone 17 Pro or Pro Max to run the 27B model.”smallerize
“Doing some naive math, the F16 filesize is ~53.8 GB … the 1‑bit version is ~3.8 GB, about 7 % of the original size.”Catloafdev
3 Performance trade‑offs & reasoning‑loop issues “‘The model gets stuck in a reasoning loop … much less often than 27B in my experience.’”dofm
“It fails the ‘Jabberwocky’ test.”raylad
4 Business & strategic implications “If you read to the bottom of the page, it says they’re funded by a few people, and one of them is Samsung.”trollbridge
“Apple would punish him severely unless they cleared it in advance …”CharlesW

These four themes capture the bulk of the conversation: the mathematics of ultra‑compact representations, the practical hurdles of deploying such models on phones or modest GPUs, the observed strengths and failure modes of the quantized LLMs, and the broader industry ramifications for hardware makers and open‑source strategy.


🚀 Project Ideas

Bonsai Mobile App Hub

Summary

  • A ready‑to‑install iOS/Android app bundle that lets users download and run PrismML’s 27B ternary/Bonsai models on their phone without manual dependency hell.
  • Solves the frustration expressed by “Havoc”, “smallerize”, and others who can’t get the models to work on iPhone 14/17 Pro.

Details

Key Value
Target Audience Mobile developers and power users who want on‑device AI but lack the patience to compile custom llama.cpp forks
Core Feature One‑click model download, Metal‑accelerated inference, automatic KV‑cache tuning, and UI for recipe / chat queries
Tech Stack Swift UI, Metal shader pipeline, Flask backend for model hosting, Dockerized quantization pipeline
Difficulty Medium
Monetization Revenue-ready: In‑app purchase subscription

Notes

  • HN users repeatedly asked “where can I find this app?” and complained about missing app store links – this product directly answers that.
  • Opens a discussion channel for feedback on UI tweaks and future model support, driving community engagement.

QuantEval Studio

Summary

  • A hosted platform that runs standardized benchmark suites (gsm8k, tool‑calling, reasoning loops) on any quantized LLM and visualizes trade‑offs in a dashboard.
  • Addresses the collective need for reliable, comparable performance data highlighted by comments like “I wish evaluation were discussed more clearly”.

Details

Key Value
Target Audience Researchers, product engineers, and investors evaluating quantized models for deployment
Core Feature Upload a GGUF/MLX model, auto‑run multi‑task evaluation, generate PDF reports with Pareto plots
Tech Stack FastAPI microservice, Celery workers, Plotly.js front‑end, PostgreSQL storage
Difficulty Low
Monetization Hobby

Notes

  • Directly quotes “I’m building a setup instead of taking model devs word for benchmarks” – the service removes that friction.
  • Sparks discussion about best‑practice evaluation metrics and could become a reference point for future model releases.

TernaryConverter SDK

Summary

  • A lightweight Python library that converts any dense FP16 model into a ternary/1‑bit representation with automatic scaling, packing, and export to GGUF or MLX formats.
  • Solves the pain point of “dependencies all fail on mac” and the desire for “easy conversion without compiling from source”.

Details

Key Value
Target Audience LLM engineers and hobbyists who want to experiment with 1‑bit or ternary quantization on consumer hardware
Core Feature One‑line model conversion, GPU‑accelerated packing, optional Metal/CUDA backends, CLI for batch processing
Tech Stack PyTorch, NumPy, Rust bindings for fast packing, OpenAPI spec for CLI, Docker container for reproducibility
Difficulty Medium
Monetization Revenue-ready: Commercial license for enterprise use

Notes

  • Commenters like “I spent quite sometime trying to install their tools and nothing really worked” will find this SDK instantly usable.
  • Generates community dialogue around best packing strategies (e.g., 5 trits per byte) and encourages contributions to improve efficiency.

EdgeModel Marketplace

Summary

  • A curated marketplace where developers can buy, sell, or lease pre‑optimized quantized models (e.g., 1‑bit Gemma‑4‑12B, ternary Qwen‑27B) ready for specific VRAM constraints (≤16 GB).
  • Directly responds to the recurring question “Can we fit a 100B model in 16 GB?” and the desire for practical, size‑optimized models.

Details

Key Value
Target Audience Startups, indie AI product builders, and enterprises seeking cost‑effective on‑device inference
Core Feature Searchable catalog, pricing tiers, one‑click integration SDKs for iOS, Android, and desktop, user reviews & benchmarks
Tech Stack React front‑end, GraphQL API, Stripe for payments, Dockerized model hosting, CI for automatic benchmarking
Difficulty High
Monetization Revenue-ready: Transaction fee (5% per sale)

Notes

  • Echoes HN sentiment “If you can have the smartest possible model running in 10 GB of VRAM … that would be amazing” – the marketplace makes that tangible.
  • Sparks discussion on ecosystem sustainability, licensing, and the future of edge‑AI business models.

Read Later