SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Swiftlet lets Mac users run 80B Qwen AI model using just 4.3 GB of RAM

0
·1 views

A developer has released Swiftlet, an open-source project hosted on GitHub, enabling large language models to run locally on Apple devices with minimal memory. The tool allows an 80-billion parameter Qwen model to operate on a Mac using only 4.3 GB of RAM. Even more notably, a 35-billion parameter model can reportedly be run on an iPhone. The project was shared on Hacker News, where it garnered early attention from the developer community. Swiftlet appears to use aggressive model compression or quantization techniques to achieve such low memory requirements on consumer hardware.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Seven OCR Tools With AI Compared: Web Apps, APIs, and Self-Hosted Options

A detailed comparison of seven AI-powered optical character recognition tools has been published, evaluating them across three use cases: one-off document digitization, automated pipelines, and self-hosted deployments. The guide highlights that most OCR tools perform well on clean, high-resolution documents but struggle with angled photos, tables, and handwritten text — limitations vendors rarely disclose upfront. It also raises a data privacy concern, noting that free web-based OCR tools process uploaded files on third-party servers, which poses risks for sensitive documents like contracts or ID cards. Tools reviewed include Google Vision API, Mistral OCR, and others, each assessed on accuracy, pricing, and integration effort rather than marketing claims. The comparison is aimed at developers and small teams trying to choose between cloud APIs, web tools, and self-hosted solutions like Tesseract based on their specific needs.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Open-source maintainer publicly corrects confident but flawed isolation advice within a day

A developer maintaining safari-mcp, an MCP server for AI agents on macOS, gave a user technically accurate but contextually wrong advice about session isolation across multiple agents. The maintainer confidently recommended HTTP transport as the solution for preventing agents from interfering with each other's browser tabs, crediting his own per-session isolation mechanism. However, the user pointed out within a day that the real problem lay one layer upstream — in how mcporter, their agent runner, handles MCP clients — making the transport choice irrelevant. Because mcporter either reuses a single client for all agents or creates a stateless new one per call, the maintainer's per-session isolation never engaged in either scenario. The maintainer acknowledged the error publicly, noting the core failure was verifying his own layer thoroughly while silently assuming conditions in an adjacent layer that he had never actually checked.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

LexisNexis Uses Motion Design to Visualize 173,000+ Laws Before Showing Its Solution

LexisNexis Legal's product demo for its State Net platform opens by animating the scale of the regulatory landscape, displaying counters that lock onto 173,923 pending legislative measures and 66,751 regulatory demands. The figures expand into a 3D wireframe mesh to represent the cognitive burden compliance teams face across fragmented state jurisdictions. Rather than opening with a feature walkthrough, the demo first renders the problem visually before transitioning the chaotic mesh into a clean UI dashboard. The interface then demonstrates AI-powered filtering, narrowing thousands of entries down to a specific bill like Arizona Senate Bill 1565. The approach reflects a broader design philosophy: establish the overwhelming scale of a market problem before presenting software as the structured solution.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer builds free browser-based SVG to PNG converter with no server or uploads

A developer built imgloft.com, a free SVG to PNG converter that runs entirely in the browser, after finding existing online tools slow and privacy-invasive due to server-side file uploads. The tool uses three native browser APIs — FileReader, Canvas, and Blob — with no backend, no framework, and no dependencies. Built and deployed in roughly an hour as a static site on Cloudflare Pages, it requires no user accounts or data storage. The developer chose to ship a minimal viable product quickly rather than adding features upfront, and the site began appearing in Google search results within days of launch. The target search term 'SVG to PNG converter' sees around 40,000 monthly queries, suggesting strong organic demand for such lightweight tools.

0 comments Read more at DEV Community
Swiftlet lets Mac users run 80B Qwen AI model using just 4.3 GB of RAM · ShortSingh