Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone: VALIDATE, 72/100 | Ideas Lab

A technical achievement that runs a large language model locally on phones, but lacks a clear business model or customer problem it solves.

The pain

Developers and power users want AI capabilities without cloud dependency, latency, or privacy concerns. Currently they're forced to use cloud APIs or accept slow local inference.

The gap

The gap between 'technically possible' and 'commercially viable' is massive here. Running 120 tok/s on iPhone is impressive engineering, but there's no articulated use case, pricing model, or distribution strategy. Who buys this and why? What do they do with it that they can't do today?

Build angle

If pursuing this: Start by identifying one specific use case where on-device inference creates measurable value (e.g., offline note-taking with AI, privacy-critical medical apps, gaming NPCs). Build a vertical product first, not a general-purpose tool. Prove users will pay before optimizing the model further.

Strengths

Genuine technical differentiation, 120 tok/s on-device is non-trivial and hard to replicate
Addresses real constraints: latency, privacy, offline capability
Ternary quantization is novel and could be defensible IP
Growing market demand for edge AI as regulations tighten around data

Risks

No clear customer segment or willingness-to-pay signal
Apple controls the platform and could replicate this internally
Battery drain and thermal throttling on sustained use are unaddressed
Competing against free cloud APIs (ChatGPT, Claude) with inferior model quality
Requires developer adoption, high friction, low margins
Model quality at 20B ternary may be insufficient for real applications
Hackernews showcase ≠ market validation

Fly Labs Method

Is the pain real, is there a gap, is it the right time, can one person build it.

72

Value Equation

Dream outcome and how likely it feels, against the time and effort it costs.

45

One-Person Business

Curiosity pull, identity fit, and a path from free value to paid for a solo creator.

68

Viral Frameworks

Hook strength, shareability, and how cheaply it can be tested.

55

Builder Lens

Evidence the problem exists, timing, defensibility, and a model that fits on a napkin.

62

Five lenses, one composite. How scoring works

The angle

If pursuing this: Start by identifying one specific use case where on-device inference creates measurable value (e.g., offline note-taking with AI, privacy-critical medical apps, gaming NPCs). Build a vertical product first, not a general-purpose tool. Prove users will pay before optimizing the model further.

No market research recorded for this idea yet.

SaaS founders lack a platform for finding partners and organizing joint advertising campaigns with bloggers to reduce costs by 5 times while maintaining the same reach

Other

BUILD

85FL score

Founders waste hours daily on repetitive tasks that AI could automate instantly

Founders spend excessive time on routine tasks that AI could complete in minutes, but no simple, purpose-built tool exists to handle them automatically without complex setup.

Other

BUILD

82FL score

Owners of modern Macs can't launch old games due to complex instructions involving emulators and file conversion

Other

BUILD

79FL score

All Other ideas