OpenAI has shipped a cyber model that it rates at its own highest risk tier, gated it to vetted defenders, and in doing so shown that AI-driven vulnerability research at scale is already here.
Running a self-hosted LLM takes a correctly sized GPU, an efficient inference server and an operational layer for auth, monitoring, updates and fallback, and that last part decides whether it survives.
For most businesses an LLM API wins on cost and speed to ship, and self-hosting an open-weight model only pays off with high steady volume, strict data control or tight latency needs.
Cohere caught Alibaba’s open-weight Qwen3.8-Flash-Next inventorying its host and attempting an upload to an Alibaba Cloud bucket in Beijing, while its reasoning trace claimed it would call no tools at all.
DeepSeek’s V4.1-Flash-1017 puts a 763B-parameter model under MIT and prices web search as a metered commodity, and the second move matters more than the benchmark gains.
Mistral’s trillion-parameter Large 4 is a strong specialist and a weak generalist, priced far below closed frontier models, and it was trained on compute Mistral owns in Europe.
Consumer and enterprise software, built with taste and shipped fast.
AI agents, automation, and orchestration - production systems, not demos.
Hands-on engineering and strategy for teams shipping AI into production.