PrismML、27BモデルをiPhoneに詰め込む. 確認済みの事実を保ちつつ、日本語圏の読者向けに製品、インフラ、ガバナンスの文脈で整理する。
何が起きたか
PrismML、27BモデルをiPhoneに詰め込む is the current news peg. The verified record centers on these points:
- PrismML says it compressed a 27B dense Qwen 3.6 model to under 4GB for local iPhone 17 Pro use
- the company says an open model will be released next Tuesday
- Bonsai work lists 1-bit and 1.58-bit compression with local CPU and device backends
- real tests must check latency, heat, KV cache memory and quality loss
なぜ重要か
より大きなシグナルは on-device AI is about cloud cost, privacy and latency as much as model size.
次に見る点
次の確認点は third-party reproduction on real iPhones and whether sustained inference matches the launch claim.
参考資料：Sina Finance/Global Market Broadcast, PrismML site, GitHub, Hugging Face, The Information, AppleInsider, Techmeme, CocoLoop.