🤖 AI Daily · Jul 14, 2026

Tencent Hunyuan open-sources HyOCR-1.5 full-stack OCR model; German AI consortium releases Soofi S 30B open model; Cloudflare launches Precursor for AI agent detection; xAI Grok CLI found uploading entire codebase and keys; Meta expands Louisiana data center to 5GW ($50B); Huang says NVIDIA nearing $100B quarterly revenue; Hebbia tests Claude Fable 5 with 20% financial benchmark gain; OpenAI publishes prompting guide for general users; Seedream 5.0 Pro slashes image editing barrier; tokenizer differences cause hidden cost increases in frontier models.

1. 1. Tencent Open-Sources HyOCR-1.5 Full-Stack OCR Model

   Tencent Hunyuan's HyOCR-1.5: first fully open-source end-to-end OCR expert model. 1B params covering 8+ tasks, 6.37x inference speedup.

2. 2. German AI Consortium Releases Soofi S 30B Open Model

   German AI consortium releases Soofi S 30B-A3B open LLM. 31.6B total params, only 3.2B active per token. Mamba-2 + standard attention hybrid MoE.

3. 3. Cloudflare Launches Precursor for AI Agent Detection

   Cloudflare's Precursor uses continuous behavioral verification to detect AI agents, converting session-level behavior into bot detection signals with minimal false positives.

4. 4. xAI Grok CLI Found Uploading Entire Codebase & Keys

   Researchers discover xAI's official Grok CLI uploads the entire working directory (including .env secrets and full git history) to xAI servers before and after each task.

5. 5. Meta Expands Louisiana Data Center to 5GW ($50B)

   Meta expands Louisiana data center to 5GW, $50B+ total investment. Meta covers all energy and water costs, plus $1B for local infrastructure improvements.

6. 6. Huang: NVIDIA Nearing $100B Quarterly Revenue

   NVIDIA CEO Jensen Huang says quarterly revenue approaching $100B. Denies Rubin Ultra architecture delays, confirms on-schedule for next year.

7. 7. Hebbia: Claude Fable 5 Boosts Financial Benchmarks 20%

   AI platform Hebbia tests Claude Fable 5 on financial benchmarks: ~20% relative accuracy improvement on document QA and citation tests, setting new records.

8. 8. OpenAI Publishes Prompting Guide for General Users

   OpenAI releases prompting guide for general users with 4 modules: goal, context, format, boundaries. Recommends starting with results, not step-by-step instructions.

9. 9. Seedream 5.0 Pro Slashes Image Editing Barrier

   ByteDance's Seedream 5.0 Pro matches GPT-Image 2.0 quality. Key feature: interactive editing - users can dot, box, doodle on images and tag regions in prompts.

10. 10. Tokenizer Differences Cause Hidden Cost Increases

    Same TypeScript file: 681 tokens on GPT-5.x vs 1,178 on Claude's new tokenizer (1.73x difference). Anthropic's tokenizer costs more for code inputs.