Blog

DeepSeek's first vision model bills images at the plain text rate, quietly outscores V4-Flash on six of seven text benchmarks, and squashes every image down to roughly 800 by 800 before it looks.

A plain-English guide to Qwen-UI-Agent: the benchmarks Alibaba published, the desktop gap the coverage missed, and why you cannot download or call it yet.

Qwen-UI-Agent scores 92.2% driving real Android phones and beats Claude Opus 4.8 on mobile use. Here is the benchmark gap nobody quotes, and why you still cannot run it.

OpenAI Astra solved ten open math problems for about $2,000 of tokens, then got its own training runs paused. Here is every confirmed fact, straight from OpenAI.

Shieldstral ties a 20B model on text safety at 3B. The top four guard models sit inside 1.6 F1 points, so what actually picks your guard is hosting, licence, reasons, and how many calls one message costs.

Google publishes 0.24 Wh per prompt. Mistral publishes 1.14 gCO2e. Both are honest, neither is comparable, and the number that actually moves your footprint is the one nobody publishes: how many model calls it takes you to finish one piece of work.

Mistral's 3B open-weights safety classifier ties the 20B leader on its own text-safety chart and runs on one 16GB GPU. What it does not give you is a reason, or a hosted endpoint.

SeedRealtime is ByteDance's audio-visual full-duplex model. Here is what it does, what ByteDance published, and what you can actually call today.

A hands-on read of Meta's terminal coding agent. The isolation and the audit log are the best parts, and every benchmark gain Meta showed was measured inside Meta's own harness.