PulseAugur
EN
LIVE 03:51:56

ModelBest enables 35B MoE model to run on iPhone at 20+ tokens/sec

By PulseAugur Editorial · [1 sources] ·

ModelBest has developed a technique called Lookahead Expert Scheduling that allows a 35 billion parameter mixture-of-experts model to run on an iPhone 18 Pro Max at over 20 tokens per second. This method activates approximately 3 billion parameters per token, with expert weights loading on demand from flash storage. The company has also reduced the model's file size from 19.7GB to 14.44GB using mixed-precision quantization, and the code has been open-sourced. AI

IMPACT Demonstrates feasibility of running large-scale MoE models on mobile devices, potentially enabling more powerful on-device AI applications.

RANK_REASON Open-source release of a novel technique for running large MoE models on consumer devices. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Mastodon — mastodon.social →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

ModelBest enables 35B MoE model to run on iPhone at 20+ tokens/sec

How we ranked this

Signal score
23 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
Open-source release of a novel technique for running large MoE models on consumer devices. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
model release, infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. Mastodon — mastodon.social TIER_1 English(EN) · [email protected] ·

    ModelBest says its Lookahead Expert Scheduling runs a 35B-parameter mixture-of-experts model at over 20 tokens per second on an iPhone 18 Pro Max, activating ab

    ModelBest says its Lookahead Expert Scheduling runs a 35B-parameter mixture-of-experts model at over 20 tokens per second on an iPhone 18 Pro Max, activating about 3B parameters per token. Expert weights load from flash on demand, and mixed-precision quantisation cut the weight f…

    LINKS pandaily.com/modelbest-les-lookahead-expe…

RELATED ENTITIES

RELATED TOPICS