DeepSeek open-sourced DeepInfer while Washington moved to purge DeepSeek from federal systems. Inference stack news: CNCF llm-d 1.0 (IBM, Red Hat, Google, NVIDIA) and Google TorchTPU. Research posts: NVIDIA Cosmos3-H-Dreams, Allen AI FairStress, IBM ALTK-Evolve on τ²-bench.
Today
- DeepSeek open-sources DeepInfer — In-house inference engine pitched to cut serving cost vs closed APIs. South China Morning Post
- US orders federal agencies to remove DeepSeek — Security fears and civilian-agency ban wave. Streamline Feed
- DHS ban reaches third-party federal products — Follow-on coverage on DeepSeek removal scope. Cryptopolitan
- CNCF llm-d 1.0: production open inference — IBM-backed release for cost-efficient day-2 ops. IBM Research
- llm-d 1.0 debuts at PyTorch Conference — Red Hat, IBM, Google, NVIDIA on inference as the real bill. Martin Cid Magazine
- Google open-sources TorchTPU — PyTorch on TPUs gets easier. Help Net Security
- NVIDIA Cosmos3-H-Dreams — Generative simulation for clinical surgery workflows. Hugging Face Blog
- FairStress: fairness in LLM decision-making — Allen AI eval framework write-up. Hugging Face Blog
- ALTK-Evolve tops τ²-bench on customer service — IBM learned-memory approach without a frontier model. Hugging Face Blog
- Mistral voice agent API and speech models — Low-cost voice stack announcement. Crypto Briefing
- Open artifacts #25: V4.1 Flash, US/China controls — Interconnects weekly OSS/regulatory roundup. Interconnects AI
- Playgrounds Network: Rust framework for AI agents — Open-source agent runtime review. Quasa.io
Watch: DeepSeek / China
- DeepSeek-V4.1-Flash-1017 and API pricing — NeoTeo on the latest flash tier. NeoTeo
- DeepInfer on NVIDIA and Huawei silicon — Forkast on cross-vendor serving story. Forkast News
Watch: Open-weight · agents
- Four labs and a $20M revenue gate for agent builders — Forkast on deployment ceiling narrative. Forkast News
- GoTo 27B model for Indonesian languages — Regional open-weight stack. shattered.io
- Kalshi “best LLM of the decade” odds jump — Prediction-market noise on frontier winners. tech-insider.org
Built with utilities/ai_report_build.py (24-hour lookback). Edited for publication. About the AI Report.