AI News

Alibaba released eight Qwen 3 models under the Apache 2.0 license. For European developers, self-hosting open weights is a concrete way to keep data inside the EU and avoid per-token API costs.

OpenAI's Ultrafast mode for GPT-5.6 Sol, powered by Cerebras chips, promises 14x faster inference. We examine what that speedup unlocks in production, what it costs in EUR, and what EU users should verify first.

Uber and China's Pony.ai are bringing 2,000 robotaxis to Europe, one of the biggest autonomous-vehicle fleet commitments for the region. Here's what riders, regulators, and the AI industry should watch.

Microsoft is taking SQL Server 2025, Azure SQL Hyperscale and Fabric IQ to Barcelona this September. We break down the combined SQLCon/FabCon Europe agenda — and what the EU AI Act means for the AI features on display.

Google's Gemini 3.7 Flash lands at $0.75/1M input tokens — half the previous Flash price — with DeepSWE up 16 points to 65.3%. Here are the benchmarks, EUR pricing, and the AI Act compliance picture.

A 16GB GPU like the RTX 5060 Ti draws 180W at full load, but the real monthly bill depends on idle time, model size and your EU electricity tariff. Here is the math you can measure at home.

Since 2 August, the EU AI Act applies to any UK employer hiring in EU markets — even for remote roles. Fines can reach €35 million. Here is what recruiters and applicants should know.

For months the advice was simple: want capable AI on a budget? Use DeepSeek. The official V4 Pro launch just ended that. With OpenAI Luna now at $0.20 per million input tokens, the cheapest-model crown is split three ways.

A new Cureus protocol uses ChatGPT to train novice ICU nurses in family crisis communication and de-escalation. We break down the 91.5/100 usability data and calculate what each simulation session actually costs in EUR.

DeepSeek shipped V4-Pro-0813 with no fanfare — just a version bump in its API docs. It now ranks near the top of the open-weights coding board, while output tokens cost roughly 57× less than Claude Fable 5 and 34× less than GPT-5.6 Sol.