Device intelligence is about what an agent delivers within a time and memory budget.…



This is a AI post classified by Jev as Open & local models (a model release), kept by the AI Radar because it carries real work, not commentary.
Device intelligence is about what an agent delivers within a time and memory budget. Combining Artificial Analysis’s 16K-context intelligence, end-to-end latency, peak memory and one-minute task completion scores places four of five LFM2.5 models on the joint Pareto frontier on both iPhone 17 Pro and Galaxy S26 Ultra. LFM2.5-2.6B ranks co-first with Nanbeige 3B on intelligence among 39 models tested including models 3-10x larger, with roughly 40% less memory footprint and 3x lower latency. On iPhone, that means 2.32 GB peak memory and 8.0s end-to-end latency, versus Nanbeige’s 4.03 GB and 21
Posted by Liquid AI (36.3k followers) 1 h ago · 108 likes · 4.8k views · view the original post on X. Kept by the AI Radar as Open & local models. Tools mentioned: Intelligence Benchmarking.
More AI work like this
- A 27B model just beats GPT-6 Astra & Fable 5.1 by fifty points on CommunicationBench. — @0x0SojalSec
- Respectable for a country that can't legally buy compute and had most of its AI… — @teortaxesTex
- K2-Horizon-36B-A4B scores 25 on the Artificial Analysis Intelligence Index, matching… — @IFM_AI
- Jev-style logit read on a 4B open VLM, measured: http://glance.yohei.me 👀 — @yoheinakajima
- Deepseek Harness w/ GLM 5.3 FLASH Now Has Qwen 2.1 Image Generation Plugin... We… — @Tech2Wild
- Kev looks really cool, with a HF demo you can try right now: — @victormustar
- Introducing the Pirate Face tracker - its distributed hash table (DHT) and BitTorrent… — @thepirateface
- Jev may have an open local alternative: Laya. — @davidarngar
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 39.7k posts from 5k X accounts over the last 14 days, 1.6k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-21 18:40 UTC. Full method.