NVIDIA just unveiled Apollo, a family of open AI physics models designed to speed up scientific and industrial simulation. The models cover domains like semiconductors, fluid dynamics, and climate modeling, using architectures like neural operators and transformers. Industry giants like Siemens, Synopsys, and Cadence are already integrating them to run real-time simulations in seconds.
Microsoft is doubling down on its flagship Earth-system AI model, Aurora. In partnership with the University of Cambridge, the company pledged to keep future releases and full training pipelines completely open-source, aiming to help local weather services and energy grids build tailored forecasting tools with minimal data.
Microsoft Research just introduced MMCTAgent, an open-source framework built on AutoGen to reason across long videos and massive image libraries. Instead of relying on one-shot answers, it deploys Planner and Critic agents to break down queries, trigger computer vision tools, and cross-check evidence. On the MM-Vet benchmark, it pushed GPT-4V's accuracy from 60.20% to 74.24%.
Baidu just open-sourced ERNIE-4.5-VL-28B-A3B-Thinking under an Apache 2.0 license. The model boasts 28B total parameters but activates only 3B at runtime, featuring a dynamic zoom-in mechanism to inspect visual details. Baidu claims it beats Gemini 2.5 Pro and GPT-5-High on key visual reasoning benchmarks, though independent testing is still pending.
Meta just released Omnilingual ASR, an open-source speech recognition model supporting over 1,600 languages—dwarfing Whisper's 99. With zero-shot in-context learning, developers can extend coverage to 5,400+ languages with just a few audio-text pairs. Best of all, Meta ditched its restrictive Llama license for a true Apache 2.0 release, clearing the way for unrestricted commercial use.
Weibo AI just open-sourced VibeThinker-1.5B, a 1.5B-parameter reasoning model trained for just $7,800. On benchmarks like AIME24 and AIME25, it outscored the original DeepSeek R1 despite having a fraction of the parameters. It is tailored specifically for competitive math and coding problems in English.
Manifest AI just unveiled Brumby-14B-Base, a retrained variant of Qwen3 that completely ditches the transformer’s signature attention mechanism. Instead, it uses a recurrent architecture called Power Retention, keeping per-token compute costs flat across long contexts. The team trained it on 32 H100s for just $4,000 to match baseline performance, though they admit the cost-saving trick only works by building on existing transformer weights.
PewDiePie just went all-in on local AI, building a custom mini-datacenter with 10 GPUs—including modded 48GB RTX 4090s—to run massive open-source models offline. He created a custom interface called ChatOS, orchestrated a multi-bot "council," and plans to release his own self-hostable model soon. Why it matters: It shows local, privacy-first AI compute is rapidly hitting the mainstream enthusiast space.
Krea AI just open-sourced Krea Realtime, a 14B autoregressive model that is 10x larger than existing open-source counterparts. Running on a single B200 GPU, it generates long-form video at 11 fps. Weights and a technical report are available now.
Deepseek just open-sourced an OCR system that compresses image-based text by up to 10x while retaining 97% of information. By drastically reducing vision tokens to as few as 64 per page, it lets language models process massive context without hitting compute limits.
Cisco just unveiled Project CodeGuard, an open-source framework designed to stop AI coding agents from generating vulnerable code. It integrates across the entire coding lifecycle, flagging security risks like hardcoded secrets and missing input validation before, during, and after generation. While Cisco notes human review is still essential, it aims to keep low-hanging bugs from slipping into production.
Microsoft just launched ExCyTIn-Bench, an open-source benchmark testing AI agents inside simulated security operations centers with 57 live log tables. In testing, GPT-5 with high reasoning led with a 56.2% score, but dropping its reasoning depth tanked performance by nearly 19%. Why it matters: Static Q&A tests are useless for infosec—defenders now have a reliable way to see if an AI agent can actually investigate complex threats.
Andrej Karpathy just dropped nanochat, an 8,000-line codebase that packs an entire end-to-end ChatGPT training and inference pipeline into a single repo. Running a single script on an 8xH100 node costs about $100 and yields a working, chat-ready model in just four hours, covering everything from tokenizer training to GRPO reinforcement learning. Why it matters: It turns the entire modern LLM post-training playbook into a readable, easily forkable playground for developers.
Anthropic just released Petri, an open-source auditing tool that deploys automated agents to grill AI models in multi-turn conversations and flag dangerous behaviors like deception, self-preservation, or unauthorized whistleblowing. In pilot tests across 14 frontier models, Claude Sonnet 4.5 edged out GPT-5 as the lowest-risk system. Why it matters: as models get smarter, manual red-teaming can't scale—now researchers can stress-test frontier systems in minutes.
IBM just open-sourced Granite 4, a new language model family combining Transformer attention with Mamba-2 state space architecture. The models range in size from 3 billion to 32 billion parameters. In benchmarks, the 7B hybrid model consumed just 15GB of RAM—one-sixth of its predecessor—dramatically lowering long-prompt inference costs.
