Weekly AI Roundup: Open Inference, New Flagship Models, and a Supply-Chain Scare
This week's AI news centered on five developments. DeepSeek released the inference software behind its production service under an open license, while Mistral began public testing of its largest model yet. OpenAI and Google Cloud resolved a rift, Anthropic introduced a proactive assistant named Claude Duet, and a widely used Python library for LLM proxying was compromised in a supply-chain attack.
DeepSeek published DeepInfer under Apache 2.0 for its V4.1-Flash models. It supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, with vLLM as its main foundation. At first it handles offline inference only; full serving would require additional KV-cache and prefill/decode components.
Mistral’s Large 4 public preview uses a sparse mixture-of-experts architecture: about one trillion parameters overall, with roughly 52 billion engaged per token. Training ran on 3,800 NVIDIA Grace Blackwell GPUs in European data centers; weights are slated for release by month’s end after red-teaming.
Open inference tools and new flagship models could lower costs and widen choices for developers, startups, and enterprises building AI products. Cloud partnerships may simplify procurement for large customers, while European training and planned weight releases may affect researchers and regulators weighing transparency and sovereignty. The proxy-library compromise may remind security teams that AI supply chains can expose users to broad risk, making audits and dependency controls more important.