Dec 10, 2025

• Anthropic has donated the Model Context Protocol (MCP) to the Agentic AI Foundation, a collaborative initiative under the Linux Foundation, aimed at fostering open standards in agentic AI development, supported by major tech companies like Google and Microsoft.

• Mistral Releases Devstral 2 (72.2% SWE-Bench Verified) and Vibe CLI

  • Devstral 2 is a 123B-parameter coding model achieving 72.2% on SWE-bench Verified, making it a leading open-source alternative with significant cost efficiency compared to competitors like Claude Sonnet.

• When a video codec wins an Emmy

  • The AV1 specification has been awarded a Technology & Engineering Emmy Award for its transformative role in video content delivery, showcasing the importance of open standards in the digital landscape.

• Qwen3-Omni-Flash-2025-12-01：a next-generation native multimodal large model

  • Qwen operates across multiple domains, including qwen.ai and chat.qwen.ai, indicating a robust infrastructure for AI-driven applications and services. Link to article

• New benchmark shows top LLMs struggle in real mental health care

  • MindEval is an open-source framework developed by Sword Health to rigorously assess the clinical competence of Large Language Models (LLMs) in mental health, addressing the critical need for reliable evaluation tools in AI therapy.

• Qualcomm acquires RISC-V focused Ventana Micro Systems

  • Qualcomm's acquisition of Ventana Micro Systems enhances its RISC-V CPU expertise, positioning the company to leverage open-source architecture for diverse applications in AI and IoT.

• Post-transformer inference: 224× compression of Llama-70B with improved accuracy

  • This paper presents a novel method for eliminating transformers from inference, achieving 224× compression of the Llama-70B model while improving accuracy by an average of +1.81 percentage points across classification tasks.

• Terrain Diffusion: A Diffusion-Based Successor to Perlin Noise

  • Terrain Diffusion is a groundbreaking algorithm that enhances procedural noise functions like Perlin noise by integrating diffusion models for improved realism and coherence in infinite terrain generation.

• OpenEvolve: Teaching LLMs to Discover Algorithms Through Evolution

  • OpenEvolve integrates large language models (LLMs) within an evolutionary framework to autonomously discover algorithms, enhancing traditional methods that rely on heuristics and exhaustive searches.

• Scientists create ultra fast memory using light

  • Scientists at USC ISI and UW-Madison have developed the first regenerative photonic memory, a significant advancement that enables ultra-fast, light-based computing, addressing the limitations of traditional electronic systems.

• Benchmark: Massive degradation in NVMe Random Read throughput on A100 vs H100 during Multi-GPU Model Loading

  • Significant degradation in NVMe Random Read throughput occurs on A100 systems, dropping from ~1.71 GiB/s to ~0.22 GiB/s when scaling from single-GPU to multi-GPU loading, while H100 systems show improved performance.

• Strengthening cyber resilience as AI capabilities advance

• 3 Ways NVIDIA Is Powering the Industrial Revolution

  • NVIDIA's GPU platforms have overtaken CPUs in supercomputing, enabling unprecedented advancements in AI, science, and business efficiency, as evidenced by over 85% of the TOP100 supercomputers now utilizing GPUs.

• How NVIDIA H100 GPUs on CoreWeave’s AI Cloud Platform Delivered a Record-Breaking Graph500 Run

  • NVIDIA's H100 GPUs achieved a record-breaking 410 trillion traversed edges per second (TEPS) on the Graph500 benchmark, utilizing 8,192 GPUs to process a graph with 2.2 trillion vertices and 35 trillion edges, demonstrating unprecedented performance in graph processing.

• GigaTIME: Scaling tumor microenvironment modeling using virtual population generated by multimodal AI

  • GigaTIME leverages multimodal AI to transform standard H&E pathology slides into virtual mIF images, enabling unprecedented insights into tumor microenvironments across 24 cancer types and 306 subtypes through a dataset of 300,000 images.