HiDream-O1, Hermes Agent vs OpenClaw & Claude Code in the Wild
Welcome back. Let's get into it.
Starting with some fresh image model news — fal.ai dropped HiDream-O1-Image on May 10th, and this one's architecturally interesting. It's a unified pixel-level transformer — meaning it processes raw pixels, text, and task cues all in one token space. No separate encoders, no hand-offs between modalities. One unified pipeline.
What can it actually do? Long-text layouts, poster generation, multilingual copy — the kind of stuff that usually breaks other models. It also handles subject-driven generation, so consistent faces, outfits, and IP characters across multiple images. That's a hard problem, and they're claiming it works well here.
There are four variants live right now — a Dev text-to-image with references, a Dev edit mode, a standard text-to-image with refs, and a standard edit endpoint. All on fal.ai if you want to try it.
Meanwhile, Krea's K2 model is turning heads this week. Designer Jerrod Lew used it to generate four consistent style images for a fictional brand called "Explo" — product shots, brand items — and said he was, quote, "obsessed" with how it bends and adapts a single visual style across completely different concepts. That's the thing with K2 right now — style consistency without it feeling rigid.
He then took those K2 visuals and fed them as style references directly into Seedance 2.0 prompts — and got a five-second product ad video out the other side. "Loving the results," he said. That combo — K2 for stills, Seedance 2.0 for motion, using the stills as style anchors — is a genuinely slick little pipeline.
And separately, someone named JCKHLRY tested what sounds like an unreleased Krea tool — wouldn't say what it was — but described being "genuinely shocked. It's extremely good, accurate, and fast." Whatever Krea is cooking next, people who've seen it are excited.
Now — here's the story of the week, honestly. The Hermes Agent versus OpenClaw situation. This one has layers.
First, the security problem. OpenClaw's ClawHub skill marketplace got hit by something called ClawHavoc — a supply-chain attack where bad actors uploaded hundreds, eventually over a thousand, fake skills and plugins. These weren't just useless — they were actively malicious. Trojans, Atomic Stealer — also known as AMOS — credential stealers, keyloggers. The whole package. The warning going around was blunt: "DON'T INSTALL SKILLS. MAKE YOUR OWN." If you're running OpenClaw with third-party skills installed, that's worth auditing right now.
Then — on top of that — Hermes Agent hit number one globally on OpenRouter across all AI apps. Two hundred and seventy-one billion tokens. OpenClaw was number two. The person posting this wasn't subtle about it — called OpenClaw "bloat," "TypeScript that phones home," "corporate wrappers pretending to be open source" built by nearly a thousand contributors. Strong words.
And then Teknium — who knows the space well — replied to someone claiming "Hermes with GPT 5.5 is better than OpenClaw with Opus 4.7" — and said, quote: "Ironic since OpenClaw is built by OpenAI staff now lol." Make of that what you will.
On the technical side — someone ran Hermes Agent on a DGX Spark using the step-3.5-flash-REAP-121B mixture-of-experts model — 11 billion active parameters per token, Q6_K quantization — consuming 105 gigabytes of the 121 gigabytes of unified memory available. Sixteen gig of headroom. Running at 12.21 tokens per second on streaming completions for agentic loops. That's a real benchmark on real hardware — not a vibe post.
Alright — coding agent stories, and there are some good ones this week.
First one: a developer gave Claude Code a feature to build — then went to lunch. By end of day, Claude had written 847 lines, tested them, and committed. While the developer scrolled their phone. Their coworkers were still manually debugging in the background. The setup used git checkpoints plus Claude Code, which gave the agent enough scaffolding to work autonomously without going off the rails.
Y Combinator's CEO Garry Tan had his own Claude Code moment — though more of an emergency than a flex. OpenClaw corrupted its own PATH inside Dockerfiles. He spent the morning using Claude Code to dig through the guts, find the bug, and remove it. Flying again by afternoon. That's a practical, unglamorous use case — but honestly that's the most useful kind.
There's also an interesting workflow shift from someone building OpenCode at Anomaly Co. — he's moved away from what he calls "layer-by-layer" generation — building an app like a 3D printer, one slice at a time — and toward what he's calling "progressive rendering." You start with a blurry, full-app version, then do iterative full passes that sharpen the entire shape. He's running this with GPT-5.5 and voice prompting. The analogy is good — instead of building bottom-up, you're resolving the whole thing in passes. Worth thinking about if you're building with coding agents.
And finally — a Chinese builder shared a full agentic sales stack built entirely in Claude Code Router using seven Claude Sonnet 4.6 agents. Here's how it breaks down: Scout pulls leads from LinkedIn and Apollo. Diagnoser writes outreach briefs. Builder constructs five-step sequences in Smartlead. Pitcher sends them. Checker evaluates performance. And Mobile handles iPhone replies and Calendly bookings. The whole thing runs 3.4 million tokens a day — costing $540 a month in API fees — and is closing 38 B2B clients per month at $3,000 each. Sixteen percent reply rate. A human only steps in to approve deals over $4,000 or when reply rates drop below 11 percent. That's a real business, mostly running itself.
That's your AI digest for 11 May 2026.