Krea 2 Launch, FLUX Outpainting, Fara Computer-Use Agent & Ring 2.6 Reasoning Model
Welcome back. Let's get into it.
Krea.ai just dropped early access to Krea 2 — their first foundation model built entirely from scratch. And they're rolling it out in waves. Three codes went out — K2-NF55KA, K2-TCA7ZX, and TMYVGV — each unlocking access for 80 users at a time. What makes Krea 2 interesting isn't just that it's new — it's that the whole thing is engineered around aesthetic diversity and stylistic control. Most image models converge on a house style. Krea is explicitly trying not to do that. Early access is live. Watch this space.
Meanwhile, Black Forest Labs — the team behind FLUX — just shipped FLUX Outpainting through the BFL API. And this one's a big deal for anyone doing production image work. Outpainting, if you haven't used it, lets you expand an image beyond its original borders. The problem has always been seams — that ugly edge where the new content doesn't quite match. Light breaks. Texture drift. BFL says they've solved that at the model level, not with post-processing tricks. You can expand to any aspect ratio and the image just… holds together. That's a meaningful step forward for anyone building on top of FLUX.
Staying in the image world for a second — the Draw Things app just released something called the Super Great Ultimate Upscaler LoRA. Yes, that's the actual name. It runs on Flux.2 Klein 9B, uses a four-step generation process, adds high-frequency texture detail, and strips out compression artifacts. The result? Genuinely photorealistic upscaling. If you're doing any local image work on Draw Things, this one's worth grabbing immediately.
Okay — shifting to agents. Microsoft Research just released Fara 1.5 — a computer-use agent, open-weight, available in three sizes: 4B, 9B, and 27B. It's fine-tuned from Qwen 3.5, and the way it works is — it looks at screenshots. No accessibility trees, no special browser hooks. It just sees what you see — and then it clicks, types, and scrolls. The wild part? Fara-7B — a mid-size variant — is matching much larger agents on web tasks. That's a strong efficiency story. Open-weight computer-use agents are getting very competitive, very fast.
Now — Ring 2.6. One trillion parameters. Open-source. Reasoning model. It launched this week and people are already stress-testing it hard. Developer Parul Gautam fed it a 38,000-token raw codebase — in a single prompt — and asked it to generate structured documentation, flag inconsistencies, and draft a changelog. It came back with perfect section coherence across the whole thing. And here's the kicker — she ran it on OpenRouter at literally zero cost. A trillion-parameter reasoning model. Free. That's not a typo.
Quick one — Resemble AI dropped DramaBox this week. It's an open-source voice model built for cinematic text-to-speech. We're not talking polished assistant voice here. We're talking emotions. Sighs. Grief-cracked performances. The kind of vocal texture that usually requires a recording booth and an actor. DramaBox does it without either. If you're building anything narrative — games, audiobooks, film pipelines — this is worth a serious look.
OpenRouter added Recraft V4.1 image models to the platform — and there's a whole lineup. V4.1 and V4.1 Pro for high-aesthetic general use. Vector and Pro Vector for clean SVG illustrations. Utility and Pro Utility for restrained, product-focused shots. What's getting attention is the natural photorealism and how well it handles smooth gradients — even on short prompts. No elaborate prompt engineering required.
And finally — Tencent's Hy3 preview model is doing something remarkable on OpenRouter. It hit number one in call volume after going behind a paywall — and it's holding 80% retention. Think about that. People are choosing to pay for it. Why? The pricing tells the story — its input cost is 45 times cheaper than Claude Sonnet 4. Thirty times cheaper than GPT-4.1. If the quality holds — and clearly users think it does — that's a pricing story that's going to pressure the whole market.
That's your AI digest for 15 May 2026.