🤖 AI HOT Daily · Sep 22, 2026

Xiaomi ships open-source MiMo-V2.6 (native full-modal Pro + Flash), trained through scaled reinforcement learning on the road to RSI (recursive self-improvement) — the Pro matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, scores 46 on Artificial Analysis' Intelligence Index (up from 26, best among open-weight models), and lands ~10th on Code Arena: WebDev at 1628 AutoEval (~3rd among open MIT weights, +153 over V2.5-Pro); xAI launches Grok 4.7 as its strongest coding-and-knowledge-work model at $2/M input and $6/M output tokens (same price and speed as 4.6, plus a 2x-faster double-priced variant) — tech work now frontier-grade, with Musk claiming #3 agent-coding behind Anthropic and OpenAI, an AA-Briefcase of 1657 Elo (+111), and a 46 Intelligence score trailing only Claude Opus 5 and Claude Fable 5.1; Alibaba's Qwen puts out Qwen-Image-2.1 — a 7B single checkpoint doing both image generation and editing, up to 10 reference images, a built-in prompt-enhancer LLM, diffusers/ComfyUI integration, and a no-install browser demo; Moonshot AI's Kimi Code lands official desktop clients (Kimi Code Desktop 1.0) on macOS (Apple and Intel) and Windows in sync; British Columbia sues OpenAI in California for failing to refer flagged ChatGPT activity to police before the Feb 10, 2026 Tumbler Ridge shooting that killed 8 (including five children and an educator) and wounded 27; Amazon blocks Meta's personal agent Muse from shopping on its site — claiming the access was never licensed and that Muse hides its identity while grabbing and storing user credentials, a privacy-and-security standoff; reporting to the US Congress, Nathan Lambert says China's open-weight models now lead in downloads, benchmarks, and academic adoption — leading Hugging Face downloads by ~1.6B since July 2025 (3.2B total, roughly double the US); Tomer Tunguz argues the latest wave of AI is commandeering software's if-then primitives — dedicated deciders like Jev and SemIf return in hundreds of milliseconds at 76x-209x lower cost than generative calls, and he swapped ~a quarter of his own agent's calls; Linear reworks its CI pipeline to clear the AI-coding verification bottleneck, cutting PR waits from 6+ minutes to just over 5 and roughly halving unit-test runner time; and an interview with fansub and manga scanlation groups finds them unbothered by AI — transcription/timing fell from 3-5 hours to 20-60 minutes after handing it to models, one developer built his own tool to single-handedly subtitle a two-hour live stream, while manga typesetters stick to manual work over text-embedding quality, valuing community and passion.

1. 1. Xiaomi's Open MiMo-V2.6: Full-Modal Pro and Flash

   Scaled-RL packed with an eye on RSI; the Pro rivals Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks and tops open-weight at 46 on the Intelligence Index.

2. 2. MiMo-V2.6-Pro Ranks ~10th on Code Arena: WebDev

   At 1628 AutoEval it sits ~10th overall — ~3rd among MIT-licensed open weights — climbing 153 points over V2.5-Pro's 1475.

3. 3. Grok 4.7: xAI's Strongest for Coding and Knowledge Work

   $2/M in, $6/M out; Musk claims #3 in agentic coding, and its 1657 Elo (+111) on AA-Briefcase trails only Opus 5 and Fable 5.1.

4. 4. Qwen-Image-2.1: 7B, One Checkpoint, Generate and Edit

   Up to 10 reference images plus a built-in prompt-enhancer LLM; ships in diffusers and ComfyUI with a no-install browser demo.

5. 5. Kimi Code Desktop 1.0 Arrives on macOS and Windows

   Both macOS (Apple and Intel silicon) and Windows builds ship together, downloadable at kimi.com/code.

6. 6. British Columbia Sues OpenAI Over Flagged Chat Gaps

   The province claims flagged ChatGPT activity should have been referred to police before the Feb 10 shooting in Tumbler Ridge that killed eight — five of them children — and wounded 27.

7. 7. Amazon Blocks Meta Muse's Shopping Access

   Amazon says Muse hides its identity and harvests account credentials — a privacy/security risk; the feud escalates.

8. 8. Lambert: China Now Leads the Open-Weight Race

   His congressional briefing finds China ahead in downloads, benchmarks, and academic adoption — leading HF downloads by ~1.6B since July 2025 (3.2B total, roughly double the US).

9. 9. Toggs: AI Is Coming for the If-Statement

   Dedicated deciders like Jev and SemIf answer in hundreds of milliseconds at 76x-209x less cost; Tunguz swapped ~a quarter of his own agent's calls.

10. 10. Linear Reworks CI to Beat the AI-Coding Bottleneck

    PR wait times fall from 6+ minutes to a hair over 5, with unit-test runner time roughly halved.

11. 11. Fansub and Scanlation Groups: Not Anti-AI

    Timing collapsed from 3-5 hours to 20-60 minutes; one dev single-handedly subtitled a two-hour stream with a self-built tool; manga typesetters stay manual over run-in quality.