🤖 AI Weekly Recap (Week 34)

Multi Format Support: It strips invisible Unicode artifacts (zero-width spaces, bidirectional control characters, tag characters) plus file-level metadata including hard-bound C2PA manifests, EXIF, and XMP, across PNG, JPEG, SVG, PDF, DOCX, HTML and Markdown.

MIT Licensed Agent Skill: It ships as an MIT-licensed agent skill with Python scripts, targeting Claude, Gemini's SynthID-Text, OpenAI provenance surfaces and open research watermark classes.

The Honest Caveat: The repo does not oversell itself. Unicode and metadata removal is verifiable, but statistical text-watermark removal relies on a best-effort rewrite, and Anthropic has not released a public detector, so the tool cannot prove that rewritten text defeats Claude's mark.

Runs on One Consumer GPU: At 30B parameters it would normally need about 55GB of RAM, but Meta shrunk the footprint to under 20GB using 4-bit quantization, so it runs on PCs and Macs with a single consumer graphics card.

Built for the Agent Loop: Meta pitches it for local agents, function calling, local coding, and LLM-as-a-judge evaluation, "with or without an internet connection." It also retries failed tool calls on its own.

Category Best Agentic Scores: Meta reports category-best results on MCP Atlas (75.5), SWE-Bench Pro (51.2), AIME 2026 (94.7), and Charxiv Reasoning (78.8).

A Real Machine Per Bot: Each Bot gets its own persistent cloud computer with a browser, filesystem and terminal, signs into your tools with your own credentials, and works through multi-step jobs end to end while you are away, surfacing only when something needs approval.

Bots That Hand Off to Bots: xAI ran it internally with sales bots updating CRM records and drafting follow-ups, ops bots seating new hires and processing Gmail invoices, and engineering bots reproducing UI bugs, filing tickets, then handing fixes to a debugging bot.

Bundled, Not Cheap: It is bundled into SuperGrok Heavy, Cursor Ultra ($200/month) and Cursor Teams Premium ($120/seat/month), on desktop and iOS today. It is also the first product of the SpaceX and Cursor merger, built under the codename "Sand."

All Gains, Same Base Model: GLM-5.3 shares the same base as GLM-5.2, and every improvement comes from extended post-training, using far more long-horizon task environments and a longer post-training run.

Top Ranked Open Model: It ranks first among open-source models on public benchmarks including Terminal-Bench 3.0 and Agents' Last Exam, and scored 84.5% on CyberGym.

Trained to Find Real Bugs: Working with security teams in China, Zhipu says the model found 2,436 vulnerabilities across 269 projects, some up to 40 years old, documented in a public registry. Weights are due roughly two weeks after launch, once security reviews finish.

It Drives a Real Browser: Claude Code opens the target site in Chrome, takes full-page screenshots, and inspects computed CSS values, not just the stylesheet, including inheritance, specificity overrides, and browser defaults, while clicking, hovering, and resizing through every breakpoint.

Real Fonts and Assets, Not Guesses: Before writing any code it installs the exact fonts the site uses, sets up the color system with oklch design tokens, and downloads every image, video, favicon, and OG image into organized folders.

Parallel Rebuild, Then QA: Builder agents run one per section in git worktrees, then the worktrees merge and get visually diffed against the original. The output is a production-ready Next.js codebase, not a static copy.

Fits on Your Desk: The 27B variant is designed for local and on-premise deployment, fitting into roughly 17GB of VRAM at 4-bit quantization, and community reports clock it at 206 tokens per second on a single RTX 5090 via SGLang.

Million Token Context, Native Vision: It natively handles up to 262,000 tokens and scales to one million using YaRN, and beyond text it processes images and videos including diagrams, documents, and multi-hour video.

Punching at Frontier Level: Benchmarks show the 27B model edging out Opus 4.6 Max on LiveCodeBench and instruction following, with 73.0 on Terminal-Bench.