Hermes Agent Explosion, Cursor's Composer 2.5 & Open-Source Video Models
Welcome back. Let's get into it.
NVIDIA dropped something interesting this week. It's called SANA-WM — a 2.6 billion parameter open-source world model, and the headline feature is that it's natively trained for 60-second video generation. You give it one image, some text, and a camera trajectory — and it builds out a full video with precise camera control. The architecture is doing a lot of heavy lifting here: Hybrid Linear Attention, a Dual-Branch Camera Control system, and a Two-Stage Generation Pipeline. And the kicker — it runs on a single GPU. Full paper, code, and model are out now from the NVIDIA AI Labs. That's a serious piece of kit for anyone working in generative video.
Speaking of single-GPU setups, the FluxRT pipeline got a significant update this week — now supporting the Flux.2-Klein-4B model. The update adds INT8 mode, LoRA support, something called Daydream Scope, an auto-installer, a new GUI app, and LivePortrait support. The headline benchmark? Real-time 30fps webcam stream processing at roughly 0.2 seconds of latency — on a single RTX 5090. That's not a render farm. That's one card.
And if you're watching the open-source image model space — HiDream-O1-Image landed on Hugging Face under an MIT license. It's being positioned as an alternative to Nano Banana. It handles text-to-image, image editing, and storyboard generation — and people are pointing specifically at its text rendering capabilities as a standout. MIT licensed, open weights. Worth keeping an eye on.
Now — let's talk about Hermes, because there is a *lot* happening here.
NousResearch dropped Hermes Agent v0.14.0 — they're calling it the "Foundation Release" — and the changelog is genuinely dense. SuperGrok OAuth is in. Grok 4.3 is in, with a one-million token context window. There's an OpenAI-compatible local proxy, which means tools like Codex, Aider, Cline, and Continue can all plug straight in. First-class X search via the xurl skill. Microsoft Teams support. A native Windows beta. PyPI install. Faster browser calls. LSP diagnostics post-write. And live session transfers via a slash-handoff command. This is a foundation release in the truest sense — they're laying plumbing everywhere at once.
And the community is already building on top of it fast.
Teknium — cofounder and lead engineer at NousResearch — shipped a Kanban automation upgrade. One prompt into triage — and an orchestrator agent decomposes it into subtasks, then auto-assigns those tasks to agent profiles. Each profile now has custom descriptions. It's a proper agentic workflow management system baked right into Hermes.
Then there's the baoyu-comic skill — transplanted into Hermes Agent this week. You drop in a prompt or a source document, and it generates a multi-page knowledge comic. Six styles, seven tones, seven layouts, five presets. That is genuinely one of the wilder use cases I've seen come out of an agent framework.
And here's a use case that really caught my attention. A researcher used Hermes Agent to pull Leopold Aschenbrenner's 13F filing — faster than any manual check. The agent surfaced a portfolio rotation: out of optics plays like LITE and COHR, into memory and storage names — SNDK and MU — plus a cluster of neocloud and crypto-adjacent AI infrastructure names like CRWV, IREN, APLD, RIOT, CLSK, BITF, and BTDR. And then a big basket of semi and index puts across NVDA, ORCL, AVGO, AMD, MU, TSM, ASML, INTC, and others as hedges. That's serious financial research — automated, through a local agent. The fact that this workflow exists now is kind of a big deal.
xAI also published a full setup guide for the xurl skill inside Hermes — natural-language read and write to X, including posting, search, bookmarks, and lists. Same day, xAI enabled Grok and X Premium subscriptions directly inside OpenClaw — covering chat, image and video generation, and X post search.
And on the UI side — there's an upcoming hermes-web-ui integration showing direct Grok 4.3 image and video generation in a single panel. Free through the Nous Portal — no Seedance 2.0 costs. The NousResearch repo also crossed a thousand contributors this week. That community is moving fast.
Shifting over to coding agents now —
Cursor released Composer 2.5 — their proprietary model — and the benchmark scores are turning heads. Terminal-Bench: 69.3%. SWE-Bench Multilingual: 79.8%. That puts it right next to Claude Opus 4.7, which scores 69.4% and 80.5% on the same benchmarks. Cursor says it's cheaper and optimized for long-running agent tasks. They also doubled included usage for a week as a launch perk.
And while we're on leaderboards — xAI's Grok Voice models just landed in the top five of the Artificial Analysis Big Bench Audio leaderboard. Grok Voice Think Fast 1.0 scored 97.1%. Grok Voice Agent came in at 93.3%. Both are now in the global top five. Audio models are quietly getting very good.
OpenRouter made a meaningful infrastructure move this week too. They added agentic web search and fetch to any tool-calling model on the platform. The model decides what to search, when to search, and how often — autonomously. A new provider called p0 is powering the web search side. That's a significant capability unlock for anyone building on top of OpenRouter.
On the Claude Code front — version 2.1.145 dropped with twenty CLI changes. The big one: `claude agents –json` now outputs live agent sessions as JSON, which means you can script against them. There's also a new tool for executing Bash commands and returning output — and a fix for a bypass that was auto-approving bare assignments to non-allowlisted environment variables. Security patch and power-user feature in the same release.
Warp added subagent orchestration this week. An agent creates a delegation plan — subagents coordinate via message passing — and each subagent gets its own isolated worktree. It works with Claude Code, Codex, or Warp's own agent. This is the multi-agent-in-a-terminal future arriving quietly.
And finally — a local AI stack build that's been circulating. Two RTX 3090s — picked up used for somewhere between $700 and $900 each — running Qwen 3.6 27B or Gemma 4 31B, with OpenCode or Claude Code as the agent layer, and self-hosted SearXNG for web access. The claim is that this setup matches Opus 4.5 performance — at home, on consumer hardware. Whether that benchmark holds up is debatable — but the fact that the conversation is happening tells you where the hardware curve is going.
That's your AI digest for 20 May 2026.