Qwen3.7 Max Ranks 5th on Artificial Analysis, Nearly on Par with GPT 5.4

The latest addition to the Qwen series, "Qwen3.7 Max," has ranked 5th in the scoring on Artificial Analysis, an AI model evaluation platform. It recorded a score nearly identical to GPT 5.4 (xhigh), earning a higher rating than the recently released Gemini 3.5 Flash.

The community's attention is focused on the open-weight Qwen3.7 27B/35B models. Qwen3.6 27B currently trails the Max version by 6 points, and the key question is how much of that gap this update can close. For local LLM users, improved performance in the smaller variants translates directly into practicality, so expectations are high.

Marlin-2B: Reaching State-of-the-Art Video VLM Performance with Just 2B Parameters

Marlin-2B, released by NemoStation, is a video VLM that delivers outstanding performance in structured information extraction from video despite having only 2B parameters. It is designed specifically around two questions that matter in practice for developers: "what is happening" and "when it is happening."

Specifically, it comes with two modes: scene + event captioning (with second-level timestamps) and temporal grounding via natural language queries (identifying start and end times). On rigorous benchmarks such as DREAM-1K and CaReBench, it posts scores that come within a hair of 34B-class models.

On temporal grounding, it outperforms Qwen2.5-VL-7B by +6.4 mIoU and achieves performance on par with Gemini-2.0-Flash. Because it runs on a single consumer GPU, it is expected to be adopted in edge devices and cost-conscious production environments.

Training is based on Qwen3.5-2b, using a two-stage post-training pipeline of SFT → SimPO. At inference time, two convenience methods, `caption()` and `find()`, are provided, making for a developer-friendly API design.

PixVerve: A Large-Scale Dataset and Benchmark for 100MP Ultra-High-Resolution Image Generation

Text-to-Image models have made great strides at 1K–2K resolutions, but the next frontier lies in Ultra-High-Resolution (UHR). PixVerve tackles this challenge and has built "PixVerve-95K," a dataset of 95K high-quality UHR images. Each image has at least 100 million pixels and is annotated across 7 dimensions.

Building on this dataset, the authors propose three training schemes for extending existing T2I foundation models to native 100MP generation. They have also released "PixVerve-Bench," a comprehensive benchmark for evaluating the quality of UHR images. It combines conventional metrics with multimodal-LLM-based evaluation, enabling assessment of both image quality and semantic consistency.

OpenAI Moves Into the Next Phase of "Education for Countries," Its Education Program

OpenAI has announced the next phase of "Education for Countries," its initiative to promote the use of AI in school education. The company says it will expand new partnerships, training programs for teachers, and tools for improving learning effectiveness, with the aim of improving learning outcomes globally.

Japan Topics: Messages API Added to さくらのAI Engine; Kimi-K2.6 Now Available from Claude Code

On May 20, 2026, さくらインターネット added a Messages API to "さくらのAI Engine," its AI platform. As a result, the Kimi-K2.6 model running on さくらのAI Engine can now be used directly from Anthropic's Claude Code.

Articles explaining the specific connection steps have already been published on Qiita, and the momentum of Japanese developers taking advantage of domestic AI infrastructure with low latency looks set to accelerate.

In addition, "Contextberg," a native Windows app that collects work information on your PC and provides it to AI agents as "memory" via MCP, has been released on the Microsoft Store. It enables cross-context sharing between Claude Code and Codex, and improved productivity in agent development is expected.