Swiftlet streams routed Mixture-of-Experts weights, AirLLM retook #1 on GitHub Trending with a differe, A same-day voice/TTS cluster -- livekit/agents , j

Swiftlet (HN Show #1) streams routed Mixture-of-Experts weights from disk, activating ~3B parameters per token regardless of total model size: Qwen3.6-35B-A3B runs at 2.6GB peak RAM (7-11 tok/s on an M5 Mac), and Qwen3-Next-80B-A3B runs at 4.3GB peak RAM (4.5-5 tok/s on an M5 Mac, ~1 tok/s at 2.5GB RAM on an iPhone 17). The developers' own caveat: these models 'chat and write like large models but recall facts like small ones.'

AirLLM retook #1 on GitHub Trending (+1,085 stars/day, 27,437 total) with a different technique -- one-layer-at-a-time GPU streaming rather than expert streaming. Its newest capability (added July 2026): Kimi K3 (2.8T parameters) running under 4GB. Two independent memory-reduction architectures for on-device inference converged on the same day, independent of quantization format.

A same-day voice/TTS cluster -- livekit/agents (+148/day, 12,091 total, realtime voice-agent framework), jamiepine/voicebox (+412/day, 48,826 total, local-first voice studio), and two Hugging Face speech models -- lands exactly on this vault's own two-week falsification deadline for a 'one real signal per week' speech-tooling cadence set on 07-28, resolving it in the affirmative.

zhaoxuya520/reverse-skill, the viral reverse-engineering/pentesting skill router, posts a third consecutive GitHub-API-verified growth day: 16,274 total stars, up from 14,080 a day earlier (~2,194 gained in 24 hours) -- sustained growth, though a second comparable skill router is still needed to confirm 'genuine vertical' status.

One day after ChatGPT hit its highest-ever recorded Play Store position (#2) and Perplexity made its first appearance in the captured top-10 (#9), today's 38-item capture contains zero AI-chat apps at all -- the fastest and sharpest reversal this pipeline's five-week rank-churn thread has logged, strongly supporting the standing 'tail-of-chart noise' reading over a genuine re-rank.

Hoplite (YC S26) launched 'effortlessly deploy cloud coding agents' -- the third same-genre cloud-agent-deployment product in five days, after AgentSky (07-30, Product Hunt) and YC's own internal tool QM (08-01). Three independently-timed teams converging on 'take your existing coding-agent setup to the cloud' within a week.

The OpenAI/Hugging Face and Anthropic agent-containment-breach stories recirculate on Tildes ~tech for roughly the fifteenth day running, with zero new information beyond what this vault logged 07-20 through 08-02.

Hugging Face Trending's top slots today are almost entirely Chinese-lab models (Kimi K3, DeepSeek-V4-Flash-0731, MiniMax-H3, GLM-5.2, multiple Qwen3.6/3.8 fine-tunes) with no US-frontier-lab open-weight entrant visible in the captured top 20.