Kimi K3: Open Frontier Intelligence
Moonshot AI announced Kimi K3, a 2.8-trillion-parameter mixture-of-experts model (896 experts, 16 active) built on its Kimi Delta Attention architecture, with native vision and a 1-million-token context window — the company claims roughly 2.5× the scaling efficiency of Kimi K2. It leads open models on agentic-coding benchmarks (Terminal Bench 2.1: 88.3) while Moonshot itself concedes it still trails Claude Fable 5 and GPT-5.6 Sol. The API is live at $3 per million input tokens and $15 output — a sharp increase from K2.6 that exactly matches Claude Sonnet pricing, a first for an open-weights lab — and full weights are promised by July 27. Simon Willison's early hands-on found strong vision capabilities and a single fixed reasoning-effort mode.
Sources & depth
NotebookLM is now Gemini Notebook
Google is retiring the NotebookLM name: the research tool — now at 30 million users and 600,000 organizations — becomes Gemini Notebook, staying a standalone product but syncing with the Gemini app and, soon, Google Search's AI Mode. The substantive addition is code execution: notebooks get a secure cloud sandbox that writes and runs code natively for data analysis, opening a new class of outputs beyond summaries and audio overviews. Rollout starts with AI Ultra and Workspace business tiers now; Pro users on the web get code execution over the coming weeks.
Sources & depth
LM Studio Bionic: the AI agent for open models
LM Studio launched Bionic, a separate desktop agent app built specifically for open models: it inspects codebases and suggests edits with inline diffs, works over PDFs and spreadsheets, and takes voice input via local transcription with Mistral's Voxtral. It runs natively local with any model you download, with optional access to hosted frontier open models (GLM 5.2, Kimi K2.7 Code) through LM Studio's cloud under a zero-data-retention commitment. It's the clearest attempt yet to give the local-model ecosystem a first-party answer to Claude- and Codex-style desktop agents.
Sources & depth
$100 AI Music Video: Claude Fable 5 vs. GPT-5.6 Sol
TryAI gave Claude Fable 5 and GPT-5.6 Sol the same song, the same brief, and budgets of $25 and $100 each, then let each model autonomously direct a music video — choosing its own generation models and editing with ffmpeg, every tool call logged, harness and transcripts open-sourced on GitHub. Fable 5 produced the subjectively better video but spent $73.65 to GPT-5.6 Sol's roughly $40; neither model ever reviewed its own output before finishing. It's vendor-published (TryAI sells access to both models), but the open harness makes it one of the more reproducible agentic-creativity comparisons around.
Sources & depth