Models & Research

AIAI-generated summary Baidu has released the ERNIE 5.5 Flash Base open-source foundation model featuring a mixture-of-experts architecture with 280 billion parameters. According to Baidu, the model competes with similarly sized open foundation models in reasoning, code, and long-context tasks.

AIAI-generated summary OpenAI says it has introduced GPT-6 Cyber, a cybersecurity model trained to find and validate exploitable vulnerabilities while staying within set scopes. OpenAI says the model scored 48.5 percent on ExploitGym and outperformed GPT-5.6 Cyber across key benchmarks.

AIAI-generated summary OpenAI says its security researchers built an autonomous pipeline that generated patches for 948 Linux kernel privilege-escalation bugs. Moonshot released weights for its Kimi K3.1 Pro model, which Modal describes as a post-training update with expanded RL for coding.

AIAI-generated summary ai.meta.com introduces GSLM2, a generative spoken language model designed to learn linguistic structure purely from speech. ai.meta.com states that the model uses SpidR 2.0 and a multi-stream SLM to improve linguistic competence and scale efficiently with data.

AIAI-generated summary The support guide explains how users can modify settings for model selection, effort level, and thinking directly from the menu next to the send button. According to the article, adjusting these options allows users to control how much reasoning Claude applies to different types of tasks and queries.

AIAI-generated summary Latent Space discussed weight-space programming experiments where frontier models edit small neural network weights directly without gradient training. Anthropic open-sourced a provable-inference system named FlashProve that verifies responses came from a specific model without revealing weights.

AIAI-generated summary OpenAI says two team members designed a working hardware prototype in under one week using GPT-6 and Astra to connect users with Codex. OpenAI is also collaborating with Mazda and Chip Ganassi Racing to test how its models can assist automotive and motorsports engineers.

AIAI-generated summary Google Developers outlines how cheap classifiers and routers can direct tasks across Gemini models and thinking levels to manage costs. The evaluation shows that starting with Gemini 3.8 Flash on low thinking solved all fifty test tasks before adding a router.

AIAI-generated summary NVIDIA released Cosmos, a platform of world foundation models and development tools designed to address physical AI challenges. NVIDIA says Cosmos Framework provides agent skills and recipes for training, while Cosmos NIM microservices support GPU-accelerated serving.

AIAI-generated summary The study by researchers from Tel Aviv University and Stanford University is available on the arXiv preprint server. The research shows that a model can acquire general predictive skills even when it is trained only on data generated within the learning process itself.

AIAI-generated summary Google Cloud makes Gemini Omni 1.2 Flash available on Google Cloud for business users and developers. Google Cloud says the model is fast, cost-effective, and follows image, video, and audio references more closely than before.

AIAI-generated summary Flexcompute has released its new Physics Twin for Space Plume Impingement simulation tool to help mission teams with complex rendezvous and proximity operations. According to the company, the software evaluates thruster exhaust risks on nearby spacecraft and reduces computational processing time compared to hybrid Direct Simulation Monte Carlo setups.

AIAI-generated summary Tencent launches the Hy4 version of its Hunyuan artificial intelligence model with added visual understanding and computer-use capabilities. The new model is available through Tencent Cloud TokenHub and is being integrated into the company productivity platforms.

AIAI-generated summary Ricursive Intelligence previews the new Ripple-65B Preview model on Hugging Face, claiming it scores the highest on coding performance on two third-party benchmarks compared to any other open model from the U.S. The release gives outside developers a way to assess technology from a well-funded startup whose main business centers on accelerating semiconductor development.

AIAI-generated summary New findings published in Nature Machine Intelligence are available, showing that some artificial intelligence models work harder to defy stereotypes. Researchers report that five reasoning models spent more reasoning tokens on tasks that conflicted with established stereotypes, indicating greater computational effort.

AIAI-generated summary Google Research released the Agile Deliberation framework to help users build subjective AI image classifiers through an interactive workflow. Google Research says the framework reduces user fatigue and anxiety while producing higher-performing models.

AIAI-generated summary Moonshot AI has released Kimi K3.1 Pro as an open-weight, natively multimodal agentic model featuring a 1-million-token context window. Moonshot AI says the model delivers substantial improvements in coding, 3D creation, financial research, factual accuracy, and other agentic capabilities.

AIAI-generated summary Tencent Holdings announced its latest artificial intelligence large language model, Hy4, on Wednesday. Tencent Holdings claims the large language model will break into the top tier of open source AI models for the first time.

AIAI-generated summary Simon Willison writes that Qwen 3.8 27B is available to run on a laptop and can interpret neural network weights. The model successfully identified classic computer games like Breakout and Pong by reading their raw floating point number weights.

AIAI-generated summary Anthropic has released Opus 5.5, reducing customers cost per task by approximately 40% compared to Opus 5 while preserving its profit margins. Anthropic says that after launching Opus 4.5 in November 2025, total quarter-over-quarter revenue growth accelerated significantly due to expanded customer adoption.