HiDream-O1, Hermes Agent Takeover & Cursor Everywhere
Welcome back. Let's get into it.
First up — a new image model just landed, and it's worth paying attention to. On May 10th, fal dropped HiDream-O1-Image onto their platform. This is a unified pixel-level transformer — meaning it processes raw pixels, text, and task cues all in a single token space. No separate encoder, no patchwork pipeline. One space, everything together.
What does that actually buy you? Two things stand out. First — long-text layouts. Posters, multilingual copy, complex typographic alignment — this model handles that stuff cleanly in a way most diffusion models fumble. Second — subject consistency. Same face, same outfit, same IP character, across totally different scenes. That's genuinely hard to do well, and HiDream-O1-Image is getting solid marks for it.
The endpoints are live on fal — there's a dev tier for text-to-image, a separate dev endpoint for image editing, and prod versions of both. Standard fal structure.
Then just two days later — May 12th — ComfyUI merged native support for HiDream-O1-Image. Commit 8e53f001. The model itself is hosted on Hugging Face under Comfy-Org. Community workflows are already circulating — PhotogenicWeekE dropped a workflow JSON, and realrebelai built out a multi-reference editing node setup posted to Civitai. One thing worth flagging from the community feedback: the editing endpoint is where this model really shines. The text-to-image side? People are calling the outputs a bit soft — almost plastic-looking. So if you're testing this, lean into the editing workflows first.
Now — over to Qwen and some genuinely impressive local inference numbers. Someone going by sudoingX ran Qwen 3.6 27B — dense model, Q4 quantized — through Hermes Agent on a single RTX 3090. Twenty-four gigs of VRAM. A card you can pick up used for around nine hundred bucks.
The task? Build a fully playable space shooter called "Octopus Invaders." One shot. No intervention.
The result: eleven files generated, 2,411 lines of code, playable on the first load — in sixteen minutes and forty-one seconds. The model was running at roughly 41 tokens per second, peaked at 21 gigs of VRAM, and used the full 262k context window with thinking mode on. Zero bugfixes. Zero manual interventions. It even ran an autonomous browser debug loop — checking console errors by itself.
For comparison — the same task on Qwen 3.5 needed a scope bugfix. So the delta between 3.5 and 3.6 here isn't just benchmarks. It's the difference between "almost worked" and "shipped."
Speaking of Hermes Agent — there's a lot moving on that platform right now.
NousResearch just dropped an early preview of Hermes Agent Computer Use — powered by trycua. The model can now control a real computer. Background mouse and keyboard — it's not taking over your screen, you keep full control — but it's watching, clicking, and acting autonomously underneath. That's a meaningful step beyond browser automation.
On top of that — Teknium added an official LINE messenger gateway to Hermes Agent. One `hermes update` command and your agent is now reachable through LINE. The docs are live at hermes-agent.nousresearch.com.
And there's a cultural moment happening around this platform too. sudoingX made a pretty bold claim — there's "zero reason to use OpenClaw in May 2026" — pointing to Hermes handling coding, video editing, marketing design, research, browser control, and terminal access all in one place. Robert Scoble's team apparently switched weeks ago.
Now — OpenClaw isn't sitting still. steipete, who calls himself the ClawFather over there, posted video proof of generation for open issues — using Codex plus a crabbox screen recording setup — and pushed real Telegram login automation as PR number 76999. There's even a Polymarket bet on whether PR 78595 gets merged. Competition between these agent ecosystems is getting real.
Switching gears to Cursor — and there's actually a few stories stacking up here.
The big one: Cursor is now inside Microsoft Teams. You mention @Cursor in a channel, and you can delegate tasks directly to agents — or just pull Cursor context into the conversation. This is a meaningful distribution play. Cursor is no longer just in your editor. It's in the place where a lot of decisions actually get made.
Then there's a smaller story that speaks to where autonomous agents are heading. Mike Coutermarsh over at PlanetScale posted that a Cursor bot — unprompted — detected an API bug through a Sentry alert and opened a PR with a fix. Nobody asked it to. It just… did it. That's the kind of thing that sounds unremarkable until you think about what it means at scale.
And one more — Haezl, who'd been doing months of chat-based development, just discovered that the Claude Code VS Code extension gives the agent direct server access. It can view files, run commands, act autonomously — no copy-pasting back and forth from a terminal. Months of the slow way, and then one day — oh, this whole time it could just do the thing.
Also worth a mention: Ryo Lu, who's a designer at Cursor with a background at Notion and Stripe, has been building features for his personal project ryOS using Cursor Cloud agents and the Cursor SDK. TV channel surfing, a new Maps app, iPod UI with Apple Music sync. It's a flex, but it's also a real signal of what's possible when someone who knows both design and tooling goes deep on agent-assisted building.
That's your AI digest for 12 May 2026.