UW's Mobile AI Factory Auto-Builds Phone-Specific Small Language Models
UW · hf · 2026-11-16
Researchers at UW's Paul G. Allen School (Evan Li, Krishna Deshpande, Wen Cheng, Shyamnath Gollakota) present Mobile AI Factory, an agentic system that automatically builds fine-tuned, quantized small language models tailored to a given task and target phone.
Given a task spec and budget (e.g. "extract drugs and diseases from a clinical note in under 1s per request"), an orchestrator LLM decides whether to tune hyperparameters, mine real data, or generate synthetic examples, while fixed rules roll back changes that hurt accuracy. The factory searches base models and quantizations from small to large, guided by real on-device measurements across inference engines and backends.
Evaluated on ten tasks across four Android phones (OnePlus 15, Galaxy S23, OnePlus 8 Pro, Redmi 9A), the resulting on-device models match or outperform larger models on unseen test sets with far less GPU time. Local SLMs keep user data private and work offline.
More from Infra
- Rumor: DeepSeek V5 in pretraining, ~2T-param sparse MoE trained entirely on Huawei Ascend, no NVIDIA — DeryaTR_ · 2026-11-16
- Microsoft email outage delayed chlorine alerts at German theme park hotel, 15 guests harmed — evilsocket · 2026-11-16
- How Lisa Su took AMD from near-bankruptcy to $1T — and she's Jensen Huang's cousin — bigaiguy · 2026-11-16
- tandem.cpp v0.3.0: 2-6x Faster Prompt Processing, DeepSeek-V4.1 GGUFs Supported — smflx · 2026-11-16
- llmfit scans your hardware and tells you exactly which local LLMs will actually run — thisdudelikesAI · 2026-11-16
- SemiAnalysis reportedly cuts 2028 AI accelerator capex forecast by 28% — firstadopter · 2026-11-16