2 min read

Chinese Firms Ship H100-Class Chips as Qwen Unveils AgentWorld "World Models"

New Models & Releases

Baidu Releases Unlimited-OCR: Baidu has launched Unlimited-OCR, a 3.3B multilingual model designed for one-shot parsing of single images, multi-page documents, and PDFs. The model features a 32K output length and is available under an MIT license on ModelScope and GitHub.

Mistral AI Launches Mistral OCR 4: Mistral AI introduced Mistral OCR 4, a high-performance model that extracts text and structural elements like bounding boxes and confidence scores. It supports 170 languages and currently ranks first on the OlmOCRBench benchmark.

Qwen Introduces AgentWorld Series: Qwen released a series of Mixture of Experts (MoE) "world models," including 35B and 397B variants, designed to simulate environments for agent training. These models emulate OS interactions, terminal environments, and software engineering tasks to improve agentic capabilities.

Qwen 3.7 Open-Weight Models Confirmed: Reports from Hugging Face discussions indicate that open-weight versions of the Qwen 3.7 family, including 27B and 122B versions, are planned for release. This will allow users to run these advanced models locally using tools like llama.cpp.

U.S. Government Pressures Meta Over AI Safety Reviews: The Trump administration is pushing Meta to voluntarily submit its AI models for security reviews, following similar agreements from OpenAI, Google, and Anthropic. Meta is currently the only major American AI developer that has not yet complied with the request.

Legion Sues U.S. Government Over AI Export Controls: Legal tech firm Legion has filed a lawsuit challenging export controls that forced Anthropic to cut off access to its Fable 5 and Mythos 5 models. The suit argues that these regulations should not apply to hosted AI models or their generated text outputs.

Hardware & Infrastructure

Chinese Companies Shipping High-End AI Chips: Seven Chinese firms, including Huawei and Alibaba, have begun shipping AI chips that compete with NVIDIA’s H100 and H200 classes. Many of these companies have recently gone public as they work to reduce reliance on Western hardware.

Research & Comparisons

Claude Vision vs. Gemini Vision Performance: A user comparison highlights that Gemini Vision currently outperforms Claude Vision in both visual recognition and general world knowledge. However, the analysis notes that specific test cases might have been present in Gemini's training data, potentially skewing results.