Attention — Oct 25, 2026
What caught my attention on Oct 25 (6 items).
LLM Battle 11 - Qwen vs Swift
NoPlaceLikeLocalhost (Article:07:18) - Published 2026-10-20 18:00 - Discovered 2026-10-25 07:18
In this video, the host tests “Swift,” a variant of Qwen 3.8 that claims to be faster by reducing the model’s thinking output without sacrificing quality. The comparison is against Das Lab’s “Austrian Qwen,” a quantized Qwen 3.8 27B model, using the same settings, context window, and quantization to keep the test fair. The host’s main question is whether Swift’s speed advantage comes at the cost o…
Nemotron Bolt: first look and test
NoPlaceLikeLocalhost (Article:07:23) - Published 2026-10-19 18:00 - Discovered 2026-10-25 07:23
The video reviews Nvidia’s Nemotron Bolt, a 74B-parameter mixture-of-experts model compressed to about 22 GB using aggressive quantization, allowing it to run on a 24 GB GPU with a full context window. The creator notes that Nemotron models were previously disappointing for coding, and the model card emphasizes vision-language capabilities and everyday agent tasks rather than programming. In testi…
Qwen 3.8 Flash Next vs Flash Next Coder
NoPlaceLikeLocalhost (Article:07:27) - Published 2026-10-13 18:00 - Discovered 2026-10-25 07:27
In this video, the creator compares Qwen 3.8 Flash Next with Qwen 3.8 Flash Next Coder using the Strata local LLM runtime. He first notes that Flash Next has become his favorite coding model and, after tuning Strata settings and updating the software, now runs at over 60 tokens per second on his dual-GPU server, despite bandwidth limitations from an old PCIe slot. He explains that Flash Next Coder…
Dust to Dominion - building the UI
NoPlaceLikeLocalhost (Article:07:34) - Published 2026-10-07 18:00 - Discovered 2026-10-25 07:34
This episode continues the “Dust to Dominion” spec-driven vibe-coding series, in which the developer uses a local LLM, Qwen 3.8 Flash Next running via Strata, to build a space game from scratch. After earlier sessions produced an empty window, configuration system, resource manager, audio manager, and a custom UI framework with widgets that had only been tested, the goal of this session is to crea…
Building custom MCP tools with local AI
NoPlaceLikeLocalhost (Article:07:40) - Published 2026-10-04 18:00 - Discovered 2026-10-25 07:40
The video revisits Model Context Protocol (MCP) by moving from earlier Java examples to Python and building a practical document-management MCP server called docmgr. The presenter explains that FastMCP makes it easy to expose Python functions as MCP tools, and he had an AI coding assistant generate the server in about an hour. The server wraps a DocumentStore class that can list, search, read, edi…
20261023 - We put human neurons into a computer! AI will surely follow
Pivot to AI (Podcast:18:59) - Published 2026-10-23 10:40 - Discovered 2026-10-25 18:59