LongCat 2.0 Launches as Large-Scale MoE Model With 1.6T Total and 48B Active Parameters

·1 views

LongCat has released version 2.0 of its AI language model, built on a Mixture-of-Experts (MoE) architecture. The model features a substantial 1.6 trillion total parameters while activating 48 billion parameters per inference. Details about the release were shared via the LongCat blog and picked up by the Hacker News community. MoE architectures are designed to improve efficiency by activating only a subset of parameters for each input, balancing scale with computational cost. The release positions LongCat 2.0 among a growing class of large-scale sparse models competing in the AI landscape.

Read the full story at Hacker News

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

Discussion (0)

Routine WordPress Bug Fix Uncovers 919 Paid Videos Exposed to Public

A fitness membership platform's support ticket about paying members being blocked from videos led a developer to discover a far larger security flaw. While investigating the access issue on a WordPress and WooCommerce stack, the developer found that all 919 paid videos were freely accessible to anyone with a link, logged in or not. The so-called paywall was merely a static text notice displayed above a video player that loaded unconditionally for every visitor. The flaw stemmed from WishList Member's content protection being applied to a parent page, but failing to cascade to AIOVG videos stored as a separate custom post type outside WordPress's standard hierarchy. The incident prompted a full rebuild of the platform's access layer and highlighted the importance of always testing the inverse of a reported bug symptom.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why High-Performing Dev Teams Still Slow Down: It's the System, Not the People

Engineering teams often feel sluggish despite having skilled, hardworking developers, and the root cause is rarely individual performance. According to systems-thinking principles, invisible process friction, unclear ownership, and sequential dependencies collectively drag down team output. Quality theorist W. Edwards Deming argued that 94% of organizational problems stem from the system itself, not the people within it. The Theory of Constraints, introduced by Eliyahu Goldratt, further holds that every system has one key bottleneck, and optimizing anything else is largely wasted effort. Frameworks like DORA metrics and bottleneck diagnostics are recommended over adding more standups or approval gates, which tend to compound friction rather than resolve it.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Beam Up CLI Tool Simplifies Static Site Deployments Across Multiple Providers

Beam Up is a new open-source command-line tool designed to streamline static site deployments from a developer's local machine. Built by the team at Rails Designer, it allows users to deploy with a single command after a straightforward configuration setup. The tool supports multiple hosting providers including AWS S3, Netlify, Hetzner, Seal Static, and Statichost, as well as SFTP. Beam Up is compatible with popular static site generators such as Jekyll, Middleman, and Bridgetown, and can also be integrated directly into Ruby scripts. It is available as a Ruby gem and defaults to Seal Static when no configuration is provided, automatically creating an account on first deploy.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why 85% of Enterprise AI Projects Fail — and What a 3-Minute Tool Reveals

A Chinese HR professional built a recruitment tool in three minutes that screened six candidates from 310 résumés, going viral with 300,000 views — while the same company's 2.75-million-RMB professional AI system, developed over six months, was scrapped after failing to deliver. Research from Gartner, MIT, and McKinsey indicates that between 85% and 95% of enterprise AI projects either fail to meet expectations or never reach production. Analysts and case studies point to a common pattern: AI adoption is driven by top-down leadership mandates rather than genuine operational needs, turning projects into status symbols rather than practical solutions. DingTalk's internal ONE project similarly collapsed under conflicting goals and organizational pressure, despite having no fundamental technology shortcomings. The core lesson is that enterprise AI struggles not because the technology is inadequate, but because projects are disconnected from the frontline workers and specific problems they are meant to solve.

0 comments Read more at DEV Community