Daily Brief
2026-09-23 · Morning · 10 items
-
Apple Sets September 22 Launch for New Mac mini and Mac Studio
Mac mini debuts M6 and M5 Pro chips, with up to 4x higher AI performance.Apple：Newsroom（RSS）
-
Hugging Face Transformers Can Now Load GGUF Models Directly
A single from_pretrained call reuses ggml's Metal kernels for local inference.Hugging Face：Blog（RSS）
-
Meta's Muse AI hit by 0-day; Amazon starts blocking it
Any local app or terminal command can steal a user's token and fully take over the agent.Hacker News 热门（buzzing.cc 中文翻译）
-
Arena opens voting for GPT-6 Sol and GPT-6 Luna
User votes will decide where both models land on Arena's public leaderboard.X：Arena (@arena)
-
Artificial Analysis reviews GPT-6 Sol and Luna: cost halved
Sol costs $2 per million input tokens and $10 output, about half of GPT-5.6's same-named models.Artificial Analysis 完整文章（网页）
-
OpenAI launches GPT-6 Sol and Luna at half the price of its predecessor
Both models beat GPT-5.6's best AutomationBench score at a lower cost per task.X：OpenRouter (@OpenRouter)
-
OpenRouter benchmarks Jev 1.13 vs Claude Opus 5 on Banking77
Jev trails Opus by 3.3 points in accuracy, but cuts latency to roughly 1/13 and costs just $0.42 per 1,000 requests.OpenRouter：Announcements（RSS）
-
Qwen-Image-2.1 released open-source, tops image-editing leaderboard
It scores 1367 on the Image Edit Arena, ranking #1 open-source and #16 overall, just 3 points shy of 15th place.X：Arena (@arena)
-
OpenRouter launches Batch API, cutting batch inference pricing in half
Requests finish asynchronously within 24 hours at 50% or less of standard per-token rates, covering 70+ models today.OpenRouter：Announcements（RSS）
-
LiteParse 2.14.6 speeds up text extraction by 20-25%
A mimalloc-tuned PDFium fork cuts extraction to 2.76ms per page.LlamaIndex：产品、工程与评测