The week in AI

It was a quiet week. Ant Group's inclusionAI team published Ring-2.6-1T under the MIT licence on 14 May. Two days earlier, on 12 May, Krea launched Krea 2, its first image foundation model trained from scratch. Anthropic and Shanghai AI Laboratory also shipped smaller releases.

The effort levels set how much reasoning the model does before it answers. All the scores Ant published are for xhigh, the higher setting. Ant reports 66.18 on ARC-AGI-V2 and 95.83 on AIME 26. It also reports 88.27 on GPQA Diamond. These are company-reported results from the model card, and the input has no independent evaluation of them.

Ant says it trained the model with asynchronous reinforcement learning (RL). In asynchronous RL, the model keeps generating practice attempts while training updates run separately, so neither process waits on the other. Ant credits a method called IcePop with keeping training stable at trillion-parameter scale, and it also cites a stick-breaking algorithm. The input doesn't say what either technique does in detail. The paper on arXiv is the place to check before explaining them on stage.

Ant also lists upgrades to how the model carries out agent tasks. The weights are on Hugging Face under the MIT licence, which permits commercial use and modification. The Hugging Face card calls Ring a reasoning sibling, which places it in a model family, but the input doesn't name the other members.

Krea says Krea 2 targets what it calls "aesthetic collapse". That is the tendency of image models to drift toward one polished house look, whatever the prompt asks for. Krea's earlier products were built on models it had not trained itself.

The release includes style transfer and moodboards, which let users steer output toward a set of reference images. It also supports LoRAs, the small add-on weight files that adapt a model to a particular style without retraining the whole model. The input gives no benchmark scores or user numbers for the launch, so the claim about style diversity is Krea's own description.

• Claude Platform on AWS became available on 11 May and gives AWS customers the full Claude API and Anthropic's Managed Agents. AWS handles authentication and billing, and the spending counts toward customers' existing AWS commitments.
• Intern-S2-Preview is a 35-billion-parameter science model from Shanghai AI Laboratory's InternLM team, released on 15 May. It continues training from Qwen3.5. The lab reports that it matches its trillion-parameter Intern-S1-Pro on core science tasks and that it can generate crystal structures for materials.