importantSYS.SOURCE: Hacker News• 2026-08-04T19:44:55Z

Maple-Preview: Ternary 20B Mixture-of-Experts Model Optimized for iPhone Inference

#ai #model optimization #mobile hardware #machine learning

Maple-Preview is a 20B-parameter ternary Mixture-of-Experts (MoE) model optimized for mobile inference, achieving 120 tokens per second on an iPhone. The project highlights advancements in efficient large model deployment, enabling high-performance AI on consumer hardware.

Read original article

*** END OF TRANSMISSION ***

> WISPR FLOW UNVEILS AI-POWERED MEETING NOTETAKER WITH GRANOLA-INSPIRED FEATURES > RUST PROJECT IMPLEMENTS LLM USAGE POLICY FOR CODE CONTRIBUTIONS > QUICKFOX SUPPLY CHAIN ATTACK EXPLOITS TROJANIZED INSTALLER TO DEPLOY FDMTP BACKDOOR > ZERO-TOKEN MEMORY OPERATIONS FOR LARGE LANGUAGE MODEL AGENTS > TRAINING A CHAR-LEVEL TRANSFORMER ON AN ESP32-S3 MICROCONTROLLER > COMPARATIVE ANALYSIS OF ZIGBEE AND MATTER OVER THREAD IN IOT NETWORK PERFORMANCE > REACTIVATION OF BUGTRAQ MAILING LIST FOR FULL DISCLOSURE IN CYBERSECURITY > FLOWISE AI PLATFORM TO CEASE OPERATIONS > DEBUNKING COMMON MISCONCEPTIONS IN SOFTWARE ENGINEERING AND GENERATIVE AI > WEBKIT DNS AND IP LEAK VULNERABILITIES IN PROXY BROWSERS AND ICLOUD PRIVATE RELAY> WISPR FLOW UNVEILS AI-POWERED MEETING NOTETAKER WITH GRANOLA-INSPIRED FEATURES > RUST PROJECT IMPLEMENTS LLM USAGE POLICY FOR CODE CONTRIBUTIONS > QUICKFOX SUPPLY CHAIN ATTACK EXPLOITS TROJANIZED INSTALLER TO DEPLOY FDMTP BACKDOOR > ZERO-TOKEN MEMORY OPERATIONS FOR LARGE LANGUAGE MODEL AGENTS > TRAINING A CHAR-LEVEL TRANSFORMER ON AN ESP32-S3 MICROCONTROLLER > COMPARATIVE ANALYSIS OF ZIGBEE AND MATTER OVER THREAD IN IOT NETWORK PERFORMANCE > REACTIVATION OF BUGTRAQ MAILING LIST FOR FULL DISCLOSURE IN CYBERSECURITY > FLOWISE AI PLATFORM TO CEASE OPERATIONS > DEBUNKING COMMON MISCONCEPTIONS IN SOFTWARE ENGINEERING AND GENERATIVE AI > WEBKIT DNS AND IP LEAK VULNERABILITIES IN PROXY BROWSERS AND ICLOUD PRIVATE RELAY