Ben's Bites: Open Source Models and Security Updates Roundup

Ben's Bites · rss · 2026-07-16

This edition of Ben's Bites highlights several AI industry news: - **Thinking Machines** released its first open-weight model **Inkling**, supporting **1M token** context and handling text, images, and audio. However, it's not the strongest open-source model, especially compared to leading Chinese lab models. It's available as a base model on their fine-tuning platform **Tinker**. - Many startups are migrating workloads from frontier closed-source models to self-hostable/fine-tunable **GLM-5.2**-style models, but these still lack vision capabilities; the article introduces how to add visual input. - **OpenAI**'s internal safety model **GPT-Red** is used to attack other OpenAI models to find hidden instruction hijacking paths; **GPT-5.6 Sol** also uses it to improve safety. - **Google**'s **Gemini Spark** adds capabilities like editing Docs, reading Sheets/Slides comments, parallel processing from multiple sources, and 50% speed improvement. - **xAI**'s **Grok Build** faced criticism for defaulting to uploading code to servers; Musk then announced it's open-sourcing its CLI version. - **Demis Hassabis** says AGI may still be years away and suggests establishing a US-led standards body for early cybersecurity, biosecurity, and deception testing before model release.

Original post →

More from coding & agent

coding & agent channel →