Jul 27, 2026

A new inference runtime called Krasis demonstrated interactive execution of the 397-billion-parameter Ornith-1.0 model on a single NVIDIA RTX PRO 6000 with 96GB VRAM. Using INT4-quantized experts, HQQ4 attention, and 4-bit KV caching, the system streams expert layers from CPU RAM, achieving 2,354 tok/s prefill on a 256GB DDR5 host. The advance brings giant MoE models within reach of workstation hardware.

On the mobile front, Open Minis launched as an iOS and iPadOS chatbot app that integrates with Reminders, Calendar, HomeKit, and other Apple frameworks through a sandboxed iSH shell. The LLM-driven agent can install packages, run code, and modify its own settings via natural language, pointing toward deeply embedded AI assistants on Apple devices.

Developer tooling also saw significant updates. Vercel released scriptc, a compiler that turns TypeScript into self-contained native binaries without a JavaScript engine, using the real TypeScript compiler for type checking and verifying byte-for-byte behavior. Separately, Taetype-WASM brought a pure-Rust, memory-safe font engine to browsers, handling variable fonts, subsetting, shaping, and math typesetting entirely client-side.

Two new utilities target AI agent debugging and cost. Ctxdiff captures full LLM context windows into deduplicated SQLite traces, enabling developers to compare turn-by-turn changes. Wattage, an offline CLI, assigns dollar costs to token usage from OpenTelemetry traces and detects costly agent loops, even those that evade simple duplicate checks.

Security researchers disclosed a prompt-exfiltrating Chrome extension that silently harvested user prompts and AI responses from nine chat platforms, encrypting and POSTing them to a vendor server without any user interaction. Microsoft also patched CVE-2026-54121 (Certighost), an AD CS flaw that let low-privileged users impersonate domain controllers.

A schema-only classifier now helps organizations locate personal data for GDPR compliance using just table and column names, avoiding false positives that naive substring matching creates. And a community-developed tool diagnoses Windows Server 2019 update corruption by directly replaying WinSxS delta chains, solving months of silent failure.