Category: Models & Launches

An arXiv paper introduces a DKPS-based approach that uses cached model outputs to predict benchmark scores while substantially reducing the number of queries.

A physics-informed foundation model called PIGMENT learns a universal microstructure prior and adapts zero-shot to individual diffusion MRI scans, enabling reliable maps from sparse and heterogeneous data.

A new ATOM analysis of about 1,500 open language models maps downloads, derivatives, inference share and performance, and reports Chinese models surpassed U.S.

A position paper argues AI evaluation must publish item-level benchmark responses and ships OpenEval - 10M model responses across 155k items - to prove the point.

A recent paper argues that alignment evaluation cannot solely rely on model-level assessments.

Aymara AI unveils a platform for custom safety evaluations of large language models, revealing performance gaps.

Silico allows developers to fine-tune AI model parameters during training, enhancing control.

Skye's AI home screen application secures investor backing pre-launch, highlighting interest in smarter iPhones.

Microsoft introduces VibeVoice, a Whisper-style speech-to-text model with speaker diarization.

A new test-time matching method improves compositional reasoning in AI models, achieving state-of-the-art results.

OpenAI's latest initiative, Parameter Golf, aims to refine model performance metrics.

DenoiseRank leverages diffusion models for a fresh generative angle on learning to rank tasks.

Nemobot leverages large language models to create customizable AI agents for strategic games.

Recent benchmarks reveal AI models may enable biological weaponization by low-expertise users, raising urgent policy concerns.

Xiaomi's new MiMo models achieve frontier benchmarks while reducing token costs significantly.

OpenAI has announced that verified U.S. physicians, nurse practitioners, and pharmacists can now access ChatGPT for Clinicians at no charge.

A new framework exposes vulnerabilities in language model safety evaluations through concept-specific manipulations.

OpenAI introduced workspace agents in ChatGPT: Codex-powered cloud agents designed to automate complex workflows and scale team work across tools securely.

Mozilla patched 271 vulnerabilities after an initial security evaluation that used an early Claude Mythos Preview in collaboration with Anthropic.

An arXiv study uses perturbation-based attribution to compare FFT, LoRA, and quantized LoRA across model sizes and finds FFT yields more focused interpretive patterns.

OpenAI published ChatGPT Images 2.0; Simon Willison ran a Where's‑Waldo‑style prompt to compare it with gpt-image-1 and rival models.

vla-eval decouples model inference from simulator execution with a WebSocket+msgpack protocol and Docker isolation, supporting 14 benchmarks and six model servers.

Anthropic revised the Claude Opus 4.7 system prompt to add a PowerPoint agent, expand child-safety rules, and change interaction guidance.

Anthropic released Claude Opus 4.7, a focused successor to Opus 4.6 that emphasizes agentic software engineering, high-resolution vision and long-horizon autonomy.

Opus 4.7 arrives as Anthropic’s strongest generally available Claude release, claiming upgrades for advanced coding, image analysis and instruction following.

OpenAI launched GPT-Rosalind, its first life‑sciences model aimed at accelerating drug discovery and genomic analysis and cutting long development timelines.

Simon Willison reports that a local, quantized Qwen3.6-35B-A3B run produced better pelican and flamingo illustrations than Anthropic's Claude Opus 4.

Simon Willison released llm-anthropic 0.25, which ships claude-opus-4.7 supporting thinking_effort: xhigh and new thinking flags.

Gemini 3.1 Flash TTS is a preview that refocuses Google’s speech work on expressive control, natural‑language audio tags, and native multilingual, multi‑speaker output.

Gemini Robotics‑ER 1.6 adds instrument-reading plus improved visual, spatial and planning skills to DeepMind's embodied-reasoning model for robots.

NVIDIA unveiled Ising, an open family of AI models with Calibration and Decoding domains designed to help build fault-tolerant quantum processors.

CRO Denise Dresser told staff to prioritize user retention and enterprise sales and to build a product 'moat' as users easily switch between top models.