9 articles
Anthropic's flagship model sets new benchmarks across reasoning, coding, and extended context — here's what changes in practice.
We built a research pipeline that synthesises hundreds of AI sources daily. Now we're opening it up. Here's what to expect.
Mistral Small 4 ships under Apache 2.0 with 256k context, 40% lower latency, and a unified architecture for instruct, reasoning, multimodal, and coding tasks.
NVIDIA's annual GPU conference delivered Blackwell Ultra with 2.5x inference throughput, a rebuilt NeMo framework, and a clear signal that the company sees inference — not training — as the next bottleneck.
GPT-5.4 arrives with computer use that beats human baselines, a million-token context window, and integrated coding — here's what it means for production systems.
Google's Gemini 2.5 Pro sets new benchmarks on coding, math, and long-context reasoning — but the real story is what it tells us about where the frontier model competition is heading.
Sources say DeepSeek is preparing V4 — a 1T-parameter MoE model with 1M context and native multimodal, built on Huawei and Cambricon chips — but multiple predicted launch dates have passed.
DeepSeek releases R2 with 671B parameters in a mixture-of-experts architecture under an open licence, posting benchmark scores within striking distance of the best closed models — and the implications ripple far beyond the leaderboard.
Mistral releases Large 3, a 123B parameter model that outperforms GPT-4o on several benchmarks — and signals that Europe's AI sovereignty ambitions now have technical substance behind them.