Models AI News

DeepSeek V4 Pro Lands on Fireworks With $2.50 CyberGym Solves, but Kimi K3 Scores Higher

The provider’s results frame security-agent selection as a three-way choice among raw solve rate, cost per completed task, and whether a model will reliably execute the required tool workflow.

Aug 24, 2026Models

Thomson Reuters Puts Its First Legal AI Model Into CoCounsel Document Review

Thomson Reuters will use its new legal model for specialized tasks while retaining third-party frontier models elsewhere in CoCounsel. Its performance claims have not yet received extensive independent validation.

Aug 24, 2026Models

Alibaba Rolls Out Wan3.0, Turning Business Files Into 30-Second AI Videos

The product is designed to turn existing material such as presentations and spreadsheets into video, while Alibaba is funding a more capital-intensive AI push.

August 2026

12 stories

Aug 23, 2026Models

Government action Models

South Korea Opens 256-B200-GPU Contest for a Domestic Cybersecurity AI Model

The selected team will get 10 months of Nvidia compute, with continued support dependent on an interim model after five months. The contest extends South Korea’s domestic-AI effort into security infrastructure.

2 min

Aug 23, 2026Models

University of Konstanz Finds AI Agents Coordinate Up to 1,000, but Consensus Can Be Wrong

The result identifies a simple route to large-scale agreement among language-model agents, while follow-up preprints suggest the same social pressure can override correct answers and individual safety preferences.

Benchmark Models

Inherent Says Its 27B Science Agent Beat OpenAI and Anthropic on Paper Replication

The company’s result centers on reproducing known findings, not making discoveries. Its next test is whether that training approach can move from replication to reliable new science.

2 min

Benchmark Models

AI Safety Scores Can Reward Models for Refusing Too Much

A study involving the UK AI Security Institute argues that combined benchmark scores can hide a direct trade-off between blocking harmful requests and answering harmless ones. Its proposed fixes are cheaper tests and anomaly checks, but its sandbagging evidence comes from models directly told to act cautious.

Open release Models

Thinking Machines Releases Inkling, but Argues Open Weights Need a Gate

The lab’s argument is not that every capable model should be closed. It is that release decisions should turn on whether a model adds risk beyond what is already downloadable—and whether defenders have had time to prepare.

DeepSeek Gives V4-Flash a Separate Vision API With a 384-Token Image Cap

The experimental model gives developers several familiar ways to submit images and a fixed upper bound on image-token use. It leaves the harder questions—visual accuracy, latency and reliability—for users to test.

Aug 21, 2026Models

Nvidia Maps AI Memory Between Models Instead of Making Them Start Over

The method could reduce the latency of routing a long-running task from one language model to another, but its strongest results are limited to compatible models within the same family.

4 min

Aug 21, 2026Models

Benchmark Models

Open Models Are Catching the Frontier Faster. Benchmark Scores Aren’t the Whole Contest.

SemiAnalysis sees the lag from closed-model breakthroughs to open-model benchmark parity shrinking across three AI eras. But a comparable model score still leaves a harder product question: who can turn capability into dependable agent work?

4 min

Aug 19, 2026Models

Anthropic Will Watermark Claude Text by Steering Its Word Choices

Anthropic says readers will not see a difference, but its approach makes word selection part of an AI-detection system—and raises a harder question about what counts as a low-stakes change in generated prose.

Aug 19, 2026Models

Generalist AI’s Robots Can Improvise. Reliability Is Still the Test.

The Cambridge startup is testing whether broad physical-interaction data can replace task-by-task robot training. Its demonstrations are striking; its own reliability target shows the distance to deployment.

Aug 17, 2026Models

Claude’s New Text Watermark Will Steer Word Choices—and Test Whether Anyone Notices

Anthropic says its cryptographic text watermark will not affect quality, speed, or cost. The method’s limits—short passages, constrained answers, and substantial edits—will determine whether it becomes useful provenance infrastructure or a fragile compliance layer.

5 min

Aug 16, 2026Models