MoE vs Edge AI: How a 35B Model Runs on an iPhone With Under 3GB RAM

·1 views

A demo by a project called Edge0 showed a 35B-parameter language model running on an iPhone using just 1–2.5 GB of peak memory, sparking interest in how this is possible. The key technique is Mixture of Experts (MoE), a neural network architecture where multiple smaller sub-networks called 'experts' exist within a model, but only a few are activated for any given input token. This sparse activation means a model can have a large total parameter count while using far fewer parameters per computation, keeping active memory usage low. In Edge0's approach, the full set of expert weights is stored on the device's storage and only the required experts are loaded into RAM as needed. MoE is an architectural design choice, distinct from 'edge computing,' which refers to where computation runs — the two concepts are related in practice but solve fundamentally different problems.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Opinion: Are AI Companies Overhyping Replacement Fears to Sell Products?

A Hacker News user has argued that AI companies are deliberately stoking fears of human replacement to market their products. The author contends that while AI can perform reasoning tasks, it cannot replicate the full depth of human knowledge, most of which is never recorded or shared as data. Humans are said to possess unique capabilities such as cross-domain thinking, intuition, and inspiration that machines cannot replicate. The post suggests that AI systems are fundamentally limited because they only train on a small fraction of human experience and knowledge. The opinion piece, which attracted minimal engagement, invites the community to share their views on whether AI displacement fears are being manufactured or are genuinely warranted.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Thai AI Models Outperform Global Rivals on Local Tasks but Lag in Reasoning

A September 2026 analysis by Nokka on DEV Community examines where Thai-developed AI models hold advantages over global frontier models and where they fall short. Thai models perform better on tasks like handwriting recognition, Thai legal citation, Isan dialect speech, and culturally nuanced phrasing, largely because they were trained on locally relevant data. However, global models significantly outperform Thai counterparts in complex multi-step reasoning, coding, and broad general knowledge, partly due to their much larger parameter counts ranging into the hundreds of billions. The parameter gap reflects a structural economic constraint, as training large models requires tens of thousands of GPUs that small Thai teams cannot afford. The author concludes that the most practical approach is combining both: Thai models for language- and culture-specific tasks, and global models where raw capability is essential.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Web Cache Poisoning: How Attackers Exploit Cache Key Mismatches to Serve Malicious Content

Web cache poisoning is a security vulnerability that arises from a mismatch between how a cache identifies requests and how the origin server generates responses. Caches use a 'cache key' — often just the URL path — to decide whether a stored response can be reused, while origin servers may factor in additional request data such as headers or query parameters. An attacker can exploit this gap by sending a request with malicious input that influences the origin's response but is excluded from the cache key, causing that tainted response to be stored. Subsequent legitimate users requesting the same URL then receive the attacker-controlled response directly from the cache. Critically, the cache itself is not compromised — it behaves as designed, making the attack difficult to detect through standard cache monitoring.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How AI-Powered Analytics Is Replacing Traditional Business Intelligence in 2026

Businesses have long relied on manual dashboards and periodic reports to interpret data, but this approach struggles with growing data volumes and slow insight delivery. AI-powered analytics replaces that pipeline by combining machine learning, natural language processing, and automation to analyze data and generate insights without manual intervention. Unlike traditional tools that only describe past events, AI analytics can explain why something happened, predict future outcomes, and recommend actions. A typical AI analytics system moves data through collection, processing, storage, and modeling stages before surfacing results as actionable business insights. The shift represents a fundamental change from analytics as a reporting function to analytics as an intelligent decision-making system.

0 comments Read more at DEV Community