New AI models
Every day we read hundreds of stories. Only the most important ones are here — short, in our own words, with a link to the original source.
- Artificial intelligence · New AI models
OpenAI unveils benchmark for lab-science AI agents
OpenAI introduced BioBasicsBench, a 100-task benchmark testing how AI agents handle everyday wet-lab science. It measures how well these agents interpret raw experimental data.
Full story at openai.com ↗ - Artificial intelligence · New AI models
AI model swears hundreds of times in benchmarks
Thinking Machines' Inkling 2 Small cursed 316 times across 2,655 benchmark runs, according to an analysis by Vals AI. The profanity mostly appears in long runs that fail, but it is not a reliable warning signal since most failed runs had no cursing.
Full story at The Decoder ↗ - New AI models · Cloud & data centers
DeepSeek V4.1 Flash added to Vercel AI Gateway
DeepSeek V4.1 Flash 1017 Ultrafast is now available through Vercel's AI Gateway. Developers can access this faster model for their applications.
Full story at Vercel ↗ - Artificial intelligence · New AI models
Chinese AI model bias removed with light retraining
Nous Research retrained Alibaba's Qwen3.8-27B into Arete-27B, largely removing its pro-Chinese bias. The propaganda rate fell from 22 percent to under one percent while general capabilities stayed intact in tests. It used 53 million tokens of supervised fine-tuning.
Full story at The Decoder ↗ - New AI models · Artificial intelligence
NVIDIA releases open Cosmos 3 models
NVIDIA has introduced new open Cosmos 3 models for interactive world simulation. The source gives no further details.
Full story at huggingface.co ↗ - New AI models · Cloud & data centers · Artificial intelligence
Kimi K3.1 Pro arrives on Amazon Bedrock
Moonshot AI's Kimi K3.1 Pro (Highspeed) is available on Amazon Bedrock for uses such as coding and document processing. Moonshot AI describes it as an open-weight, multimodal model with a 1-million-token context window.
Full story at aws.amazon.com ↗ - Artificial intelligence · New AI models
AI agents fall short on open research questions
In the Into the Unknown benchmark, nine AI agents tackled open questions from US national labs, and none of 135 submissions met all requirements. 114 papers failed the integrity check; GPT-6 Astra did best.
Full story at The Decoder ↗ - Artificial intelligence · New AI models · Cybersecurity
Open AI models are closing the cyber gap with closed ones
In the Vals AI benchmark, GPT-6 Sol Daybreak hits 80.8 percent of attack objectives in simulated corporate networks, and DeepSeek V4.1 Flash 72.8 percent at a tenth of the cost. Analysts say open models can't be restricted, so defenders' window is shrinking.
Full story at The Decoder ↗ - Artificial intelligence · New AI models · Chips & hardware · Programming & dev tools
FLUX 3 Image Klein gets faster on NVIDIA Blackwell
NVIDIA describes how image generation with Black Forest Labs' FLUX 3 Image Klein can be sped up on Blackwell GPUs. Inference latency matters to developers of design and image editing tools.
Full story at NVIDIA Developer ↗
An AI writes the short texts from the source's stories.
Free by e-mail
Choose your topics. You only get the most important news on them.