Qwen Flash-Next XXS vs S: Why S Became Our Daily Driver | Part 2

We asked three local models, run through Hermes Desktop, to audit our own StoryReserve website: Qwen3.8-27B (IQ3_S-MTP), Qwen3.8 Flash-Next IQ3_XXS and Flash-Next IQ3_S. All three noticed that our daily Romans study has 47 of 59 days live. Two called that a problem to fix. Flash-Next S called it honest, and suggested building on the fact that we publish daily. We publish one day at a time on purpose, so readers come back for the next. That is why S became our daily driver. Then the other half of the question from Part 1: how much of Flash-Next's VRAM footprint is a configuration choice? We changed one thing on IQ3_XXS, the VRAM reserve from 3072 to 8192 MiB, and measured: whole-card peak VRAM used 30.549 to 25.477 GiB, minimum free VRAM 1.311 to 6.382 GiB, expert cache 14,239 to 11,178 slots, peak PSS unchanged at 43.923 GiB. Agent-task medians 141.6 to 176.8 s with overlapping ranges; cold 100K-token document 108.7 to 121.0 s. One real task, not a universal ranking. The three audits were not run as a matched benchmark (prompts, session isolation and site snapshot unverified; deployments differ), the audits' scores are ratings of the website, and some findings conflict and were not verified. The reserve A/B is IQ3_XXS only: n = 3 per arm, sequential, not counterbalanced, on an AMD Radeon AI PRO R9700 (32 GB) with a restricted VRAM budget; 8192 MiB requested is not 8 GiB free; no game or browser was run beside it. Our hands-on report of a game, a video and an audit running at once is a subjective experience, not measured frame rate. We did not measure S against XXS speed. Chapters: 0:00 Opening 0:46 Where Part One left off 1:27 Same website, three audits 2:37 What the answer would make you do 3:26 What we did not control 4:08 The reserve A/B 5:52 In use, and the daily driver 6:36 Takeaway Part 1:    • Is the Bigger Qwen Worth It? A Real Coding...   We came to Strata through the channel No Place Like Localhost:    • Running Qwen3.8-Flash-Next in 12GB of VRAM   (credit to them and to the authors of the software; none of their measurements are used here). No public data download is available for this episode. Has your AI ever tried to fix something you deliberately designed? What task should we test next? Narrated by Mike and Teddy. Voice clone via ChatterboxTTS (not ElevenLabs). Realistic synthetic narration is used.