Inkling-Small is a mixture-of-experts model (a model that activates only a subset of its parameters per task) with 276 billion total and 12 billion active parameters, matching its larger sibling Inkling's performance at a quarter the size and running on a single NVIDIA B300 GPU in its compressed NVFP4 format.

OpenAI's new enterprise offering, Presence, aims to get AI agents into live customer service and internal workflow deployments, going further than existing Workspace Agents by targeting external use cases with OpenAI engineers assisting on complex cases.

Anthropic's Claude Opus 5 can produce complete 3D games—including a first-person shooter, kart racer, and Minecraft-style clone—generating geometry, textures, physics, and music entirely as code, and it reportedly outperforms GPT-5.6 Sol and Kimi K3 in side-by-side comparisons.

Meta AI is using a separate memory agent to maintain a structured memory bank for long-running tasks, deciding when to remind the main agent of past mistakes, which improved benchmark scores by up to 8.3 percentage points.

NVIDIA's Molt is a reinforcement learning framework built with about 8.6K lines of code that composes Ray, vLLM, and NeMo AutoModel into a single asynchronous loop, keeping agent code as ordinary Python while matching the throughput of larger Megatron-based training stacks.

Google has canceled its AI Studio app despite reportedly receiving 800,000 preorders, according to a widely discussed announcement.

An analysis of the latest open-weight model releases argues that Laguna S2.1, Inkling, and Kimi K3 demonstrate the growing practical utility of open models relative to their computational cost.

A deep-dive review of the original Microsoft GraphRAG paper plus four independent benchmark studies finds that building a knowledge graph before retrieval substantially improves answers for questions requiring connections across a whole document set, but not for simple lookup queries—and it isn't free.

A benchmark write-up finds that running the open-source Kimi K3 model on AMD's MI355X hardware delivers better performance per dollar than running it on NVIDIA's B300, positioning it among the fastest open-source LLM setups for enterprise use.

Security firm VulnCheck found that of 1,061 AI-discovered vulnerabilities in the first half of 2026, only 14 saw confirmed attacks—about 1.3%, roughly the same exploitation rate as vulnerabilities found by any method—though the time to exploitation has dropped from 120 to 80 days.

Apple's bug bounty program has been flooded with fabricated, AI-generated vulnerability reports, forcing submission caps that initially prevented an Italian startup from reporting a genuine macOS flaw worth up to $200,000.

A hands-on walkthrough builds a complete geospatial AI pipeline that trains a U-Net model alongside tools like Grounding DINO, SAM, and Mask R-CNN to extract building footprints from high-resolution NAIP aerial imagery.

Sprocket is an agentic platform aimed at streamlining both hardware and software development workflows, gaining traction on Hacker News.

Simon Willison shipped version 1.0 of condense-json, a small Python library that compresses duplicated strings in JSON data using a replacements object, which he uses to save space in SQLite logs generated by his LLM tool.

A community project tests every major AI model with the same odd prompt—generating an SVG of a frog with a Habsburg jaw—giving each model three tries per month as an informal, ongoing benchmark.