JevBench v1.3.0 -- built and run by one person

JevBench v1.3.0 -- built and run by one person, not a lab -- is the first neutral measurement of the Jev/Kev/Laya 'System One' decision-model class this vault has tracked since 09-16. Jev leads its own composite score at 74.4. Laya, flagged 09-21 as the strongest open-weights entrant on likes alone, scores 34.1% accuracy on the benchmark's hard-decision tier against Jev's 74.1% on the same tier -- attributable, per the benchmark's own methodology, to Laya's 512-token-per-question context budget truncating long inputs.

Kev gets its first neutral placement in the same benchmark: 66.7 for the 0.6B variant, inside the same competitive band as Jev, roughly consistent with Kev's own self-reported numbers (0.852 accuracy at 9B against a self-reported 0.857 for Jev) -- the closest thing to convergent evidence this model class has produced.

Xiaomi's MiMo-V2.6-Pro is MIT-licensed, 1M-context, and scores 46 on the Artificial Analysis Intelligence Index -- #1 of 114 tracked open-weight models -- performing on par with Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, ahead of Kimi K3 and Qwen3.8 Max. Its 'full modality' claim covers audio understanding per Xiaomi's own docs; this pass could not confirm speech generation, and at trillion-parameter scale it is not an on-device artifact regardless of license.

A three-person startup, Hacktron AI, used Claude Opus 5 inside OpenAI's own bug-bounty program to chain a libheif memory bug -- reached via a crafted image upload to OpenAI's Discourse forum -- into takeover of OpenAI employee ChatGPT and Codex accounts, reaching OpenAI's GitHub organization. $6,500 bounty paid, root bug patched July 27. Opus 4.8 could not produce a working exploit for the same chain; Opus 5 did, within hours of its own release.

Tim Dettmers (QLoRA/bitsandbytes) publishes specific consumer-hardware pairings for frontier-class open models: Qwen 3.8 Flash Next (125B) on a single 24GB GPU, DeepSeek V4.1 (550B) on a 128GB-memory MacBook or AMD Strix, and a 1.5-bit-quantized Qwen 3.6 running at 450 tokens/second -- a named, credentialed source making a checkable infrastructure claim rather than a benchmark-leaderboard placement.

13 of today's top-30 Google Play apps are web browsers -- a first for this pipeline's tracking, and not yet distinguishable from a scrape-window artifact. Separately, explicitly AI-named apps in the top 30 moved to 6 today (from 5/5/4 the prior three days), with only ChatGPT and Chatbot AI - Search Assistant recurring from yesterday's list -- the count keeps bouncing while the set keeps churning.

trycua/cua, logged 09-21 as 'the only tracked repo accelerating' on GitHub Trending, decelerates today from +1,018/day to +609/day -- a one-day superlative that didn't hold a second day, the kind of rolling-window noise this vault's benchmark-integrity thread has repeatedly flagged in GitHub's own star-velocity figures.

Simon Willison published the week's clearest independent explainer of the Jev/System One decision-model category the evening before this benchmark landed, naming Kev and JevBench without independently verifying either -- his read and today's neutral JevBench numbers converge on treating the class as real but still under-measured.