Reddit user converts dense models into sparse MoE without pretraining

A Reddit user reports Krea 2 produces very little variation between generations, while Qwen 2.1 yields much more diverse results. They ask how to increase variation from Krea 2.

A four-board BC-250 ex-mining cluster linked by 2.5GbE runs two instances of Qwen 3.6 35B A3B IQ4 at ~145 tok/s with 256k context, or Next Flash IQ3_XXS at ~70 tok/s with 100k context.

Filip Makraduli of Superlinked built a workflow where a scanned NDA becomes a redacted Markdown document before a coding agent reads it, using an Obsidian vault, an MCP integration, and Superlinked's inference engine. Small open models handle document processing on infrastructure the team controls.

Interview covers the seven-month effort to build Muse, described as the fastest-growing AI product since ChatGPT, and why Meta chose personal AI agents over coding. Wang, Meta's chief AI officer and Scale AI co-founder, discusses how Mark Zuckerberg and Nat Friedman shaped the product and how Zuckerberg recruited him.

The SDK release bundles Qwen Code CLI version 0.25.1, built from the same branch/ref as the SDK. It follows the prior SDK release that bundled CLI 0.25.0.

Podcast interview with Tibo Sottiaux, Head of ChatGPT & Codex, on OpenAI's culture of autonomy — including a story of shipping a change in 30 minutes that everyone assumed 17 people had reviewed, and taking production down on his third day.