How to Pick the Right Local AI Model for Your Mac Based on RAM

·1 views

A practical guide maps Apple Silicon Mac memory tiers — from 8GB to 128GB — to realistic local AI model sizes, helping users avoid overloading their systems. Because Apple Silicon uses unified memory shared by the CPU, GPU, and all running apps, available RAM shrinks quickly once macOS, browsers, and IDEs are factored in. Experts recommend keeping 25–30% of memory free to ensure models run comfortably, not just load without crashing. Smaller 2–4B models suit 8GB machines for basic tasks, while 70B models only become practical at 64GB and above. Quantization level, context length, and concurrent app usage all affect real-world performance beyond what benchmark scores suggest.

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)

Gym Tops Loggd's 2026 Habit List, But Most Users Quit Within Days

A 2026 analysis of over 6,700 habits tracked by more than 3,600 users on the Loggd app reveals that fitness-related habits — including going to the gym, exercising, and walking 10,000 steps — are the most popular, followed by hydration, reading, and meditation. However, high tracking numbers are misleading, as most habits never reach even a two-day streak, with 'go to the gym' holding the top spot alongside one of the shortest average streaks on the platform. Eight of the top ten habits appear as one-tap suggestions in Loggd's onboarding flow, meaning app design heavily influences what users choose to track. The standout exception is GitHub activity, which ranks fifth and auto-logs from a developer's commit history — making it one of the few habits that effectively tracks itself. The data highlights a significant gap between the habits people intend to build and those they actually sustain over time.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Laravel Releases Official Passkeys Stack to Enable Password-Free Authentication

Laravel has launched first-party support for passkeys through two packages: the backend library laravel/passkeys and the frontend companion @laravel/passkeys for browsers. The stack implements the WebAuthn cryptographic standard, allowing users to authenticate via biometrics, PINs, or physical security keys instead of passwords. During registration, a key pair is generated on the user's device, with only the public key stored on the server, meaning passwords and biometric data never leave the device. Developers can integrate the packages via Composer and NPM, with native support available for projects using Laravel Fortify or Jetstream. The release aims to address common password-based attack vectors such as phishing and credential stuffing by eliminating server-side password storage entirely.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

As AI Writes and Reviews Code, the Traditional Pull Request Model Faces Collapse

The rise of large language models and autonomous coding agents is fundamentally disrupting the pull request (PR) workflow that has long governed software development. Traditionally, PRs served three roles: quality control, knowledge transfer, and establishing team ownership of code. When AI both generates and reviews code, a verification paradox emerges — two models trained on similar data may reinforce each other's blind spots rather than catch meaningful errors. The loss of human narrative in code reviews also threatens institutional memory, as AI agents make probabilistic choices with no documented reasoning behind them. Industry analysts are now calling for a new paradigm based on semantic verification and property-based contract testing to address these gaps.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How QUIC and HTTP/3 Are Replacing TCP to Fix Web Performance Bottlenecks

TCP, the internet transport protocol in use for over 40 years, suffers from a flaw called Head-of-Line Blocking, where a single lost data packet stalls all subsequent data on a connection. Google began developing QUIC in 2013 as an internal project to address this limitation, and it later became the foundation of the HTTP/3 standard. Unlike TCP, QUIC runs over UDP and implements its own stream-level multiplexing, meaning a lost packet in one stream does not block others. QUIC also embeds TLS 1.3 encryption natively, eliminating a separate handshake step and reducing connection setup time compared to HTTP/2 over TCP. Today all major browsers and web servers support HTTP/3, though many developers and administrators remain unfamiliar with the technical changes and new considerations it introduces.

0 comments Read more at DEV Community