AI & design briefing for Monday, July 20, 2026

tldr

• Moonshot's Kimi K3 and Thinking Machines' Inkling headline a week of massive open-weight model drops, while Anthropic settles Fable 5 access and Alibaba previews a 2.4T Qwen.
• LocalAI turns phone video into walkable 3D at 60fps, Baidu ships a 3B OCR model for whole-document reading, and the web standards movement gets a reboot for the AI era.
• New launches include a real-time 3D scene mapper, an open-source fine-tuning SDK, and a 26M on-device model distilled from Gemini.

Top Stories

• Kimi K3: Moonshot's 2.8T open-weight model. Largest open-weight model ever, with 6.3x faster decoding at 1M-token context via Kimi Delta Attention. Demand so high Moonshot halted new subscriptions. ^[1]
• Inkling: Thinking Machines' 975B open-weight model. Sparse MoE with 41B active params, Apache 2.0, 1M-token context. Debuted as top US open-weights model on the Artificial Analysis index; integrates with the Tinker fine-tuning platform. ^[2]
• Anthropic resolves Fable 5 access. After three deadline extensions in five weeks, Fable 5 stays in Max and Team Premium at half usage caps; lower plans get a one-time $100 credit. ^[3]
• Alibaba previews Qwen3.8-Max. 2.4T parameter model available now via Alibaba's Token Plan, with full open-weight release promised. Alibaba claims it trails only Claude Fable 5. ^[4]
• Netflix: ~300 titles use generative AI. Mostly post-production; co-CEO Ted Sarandos cited 17 minutes of AI-enhanced footage in The American Experiment at half the cost. ^[5]

Deeper Cuts

• depth-anything.cpp: 3D from phone video at 60fps. LocalAI's C++ port of ByteDance's Depth Anything 3 runs entirely on local GPU, 1.3x faster than PyTorch with half the memory; smallest model is 99MB. ^[6]
• Baidu Unlimited-OCR: 3B model reads whole PDFs in one pass. Reference Sliding Window Attention keeps memory flat regardless of document length; 93.23% on OmniDocBench, runs on 8GB VRAM. MIT licensed. ^[7]
• Cline: Kimi K3 vs Fable on a real bug. Both fixed it; Fable was 3.4x faster (3.5 min, 18 tool calls), Kimi 2.3x cheaper ($0.92 vs $2.13). First open-weight model competing head-to-head with closed frontier on agentic coding. ^[8]
• Claude Code artifacts pull live data via MCP. Dashboards built once now fetch fresh per-viewer data on demand through MCP connectors; each viewer uses their own connection. ^[9]
• OpenAI CFO's 'useful intelligence per dollar' scorecard. Sarah Friar argues for grading AI by finished work, not token count; Sol beat Fable on DeepSWE with 36% lower estimated cost. ^[10]
• Dots in LLM prompts trigger hidden reasoning. Adding dots to prompts activates readable reasoning inside filler tokens, hidden from model output. ^[11]
• Designing with web standards: the AI playbook. Jeffrey Zeldman's standards movement gets a reboot for the AI era, arguing for interoperability over platform lock-in. ^[12]

Tools and Launches

• LingBot-Map. Real-time 3D scene reconstruction from video at 20 FPS, stable over 10K+ frames. Apache 2.0. ^[13]
• Tinker. Thinking Machines' open-source SDK for fine-tuning LLMs via API. ^[14]
• Cactus-Compute needle. 26M model distilled from Gemini, runs at 6000 tok/s on-device. ^[15]
• ColorSym. Free open-source color identification tool for when color alone isn't enough. ^[16]
• conversation-steganography. Hides secret messages inside LLM-generated text. ^[17]

Prompt of the day