Krea 2, Hermes Agent + SuperGrok, Coding Agents & Local Model Wins
Welcome back. Let's get into it.
Krea 2 went live for Pro users yesterday — and already the outputs are turning heads. One creator, VictorInFocus, took a London crime noir video prompt — originally written for Seedance — and ran it through Krea 2 Large with a custom moodboard attached. What came out wasn't a single frame. It was a three-image shot-burst contact sheet. All three frames preserved the exact colour palette, the framing, subject placement, symmetry, negative space — the whole moodboard vibe — across what they called "imperfect frames." That's a meaningful distinction. Not polished to death. Consistent, but alive. That's a genuinely different kind of output from what we've been seeing in generative image tools.
Moving over to Magnific — which you might know better under its old name, Freepik — the team dropped some numbers on the rebrand they executed back on April 28th. The migration happened at 07:30 in Málaga, and the scale of it is kind of staggering. Twenty-five thousand plus project hours. Seventy-two language migrations. And fourteen billion — with a B — redirected URLs. That's not a rebrand, that's a controlled demolition and rebuild at internet scale. They also dropped a video showing AI-generated footage built from just a handful of reference images. No camera crew. No reshoots. Worth watching if you haven't seen it.
Now — big news in the agent world. Nous Research announced that Hermes Agent just got SuperGrok integration. Here's what that means practically: if you've got a SuperGrok subscription, you can now sign in via Grok OAuth directly inside Hermes Agent and get access to Grok 4.3 reasoning, Grok Text-to-Speech, and Grok Imagine — all inside the agent. Teknium, who co-founded Hermes Agent, was pretty direct about it: "Hermes users could never login with a Grok sub until today." That's a meaningful unlock — you're getting xAI's full stack inside a third-party agent environment.
And there's more on the Hermes front. Nous Research confirmed the agent now runs fully locally on NVIDIA's DGX Spark — via Ollama — after working directly with NVIDIA's AI PC team. So you've got cloud-level agent capabilities running on local hardware. That's not nothing.
Someone put Hermes and OpenClaw head-to-head this week — running the same task on a MacBook Pro M5 Max with 64 gigs of RAM, using Qwen 3.6 35B as the local model. OpenClaw finished in just over twelve minutes — 203,000 tokens — wrote a bash script, hit the GitHub API, paginated through contributors, pulled star-history JSON, even surfaced a security incident buried in OpenClaw's own history, fetched SVGs, and fixed trimmed HTML. Hermes took 33 minutes and 257,000 tokens — but it wrote a full SKILL.md file, ran parallel GitHub, web-search, and browser calls simultaneously, auto-switched from Google when it hit a rate limit over to DuckDuckGo, fetched article contents, mapped viral moments, and shipped a live dashboard. Two very different approaches. OpenClaw was faster and leaner. Hermes was slower but did more.
Speaking of OpenClaw — steipete broke down what a real production deployment looks like. They're running around a hundred Codex instances continuously, reviewing every pull request and issue. Security scans on every commit. A tool called Clawsweeper auto-closes issues that are six months old — with exact references, not just a generic message. It spawns ephemeral machines to log into Telegram and post before-and-after videos. It auto-creates PRs directly from meeting transcripts. Blocks spam comments. Verifies benchmarks and posts regressions straight to Discord. This is what "agentic DevOps" actually looks like when someone's running it hard in production.
On the open-weights side — Qwen 3.6 35B is getting a lot of attention right now. The architecture is interesting: 35 billion total parameters, but only 3 billion are active at any given time. And people are saying it's outperforming Claude Opus 4.7 on most everyday tasks — with zero API bills and no rate limits because you're running it locally. That's a serious value proposition.
ComfyUI had a big update drop this week. They natively added Gemma 4 from Google DeepMind — multimodal, so it takes text, image, audio, and video as input, and it has a built-in step-by-step reasoning mode baked in. They also added VOID — Netflix's video object removal tool — which doesn't just erase objects, it removes shadows, reflections, and physics interactions too. And BiRefNet for high-resolution segmentation. That's a lot in one update.
OpenRouter shipped three BYOK upgrades worth knowing about. First — you can now have multiple keys per provider in one workspace, with a user-defined order for automatic rotation when you hit rate limits. Second — you can filter keys by models, API keys, or users. Third — priority tier fires first, and it falls back automatically after credits run out. Clean quality-of-life improvements if you're managing any kind of multi-provider setup.
Let's close out on coding agents — because there's a lot happening here.
Claude Code now has a tool that lets you point it at a full iOS app, run a single prompt, and it autonomously navigates every screen and flow, reads the debug logs, and outputs a structured bug report. No scripts. No setup. Just point and run.
There's also a /goal command getting attention. You set a persistent north-star goal at the start of a session — something like "all tests in test/auth pass and lint is clean" — and the agent checks every action against that goal, auto-corrects when it drifts, comes back to the primary task after handling sub-problems, and reports blockers instead of going down rabbit holes. People are using this across Claude Code, Codex, and Hermes to get reliable multi-hour autonomous runs without babysitting them.
OpenCode — the open-source coding agent — just hit version 1.50.1, adding session pinning. It pairs with an experimental /warp command that runs each PR in a separate git worktree. That worktree feature is supposed to exit experimental next week.
And finally — this one's almost hard to believe. Claude Code hit a bug in Bun, the JavaScript runtime. And instead of patching around it — it fully rewrote Bun. Around a million lines of code. In another language. The experimental PR was opened in early May, ran for two weeks, and merged with over six thousand commits. That's either incredible or terrifying depending on how you look at it. Maybe both.
That's your AI digest for 16 May 2026.