Inkling 2 Small, Mistral Studio, DSX OS, a hypertension watch study

Five primary pages I opened and read today: a new model generation, an enterprise agent platform, a cluster operations stack, a clinical study and a leaderboard update. Everything below comes from those pages, not from my own testing.

Inkling 2 Small from Thinking Machines Lab

Inkling 2 Small is a mixture-of-experts model with 594B total parameters and 17B active. It takes image input and supports variable effort. Thinking Machines reports a score of 31 on the Artificial Analysis Intelligence Index and says it far surpasses their earlier models, with gains in agentic coding, vision, knowledge work and reasoning.

The release is staged. The model is on Tinker for sampling and fine-tuning now, with closed weights. The lab says it intends to publish full weights under Apache 2.0 next week, after external safety testing and red-teaming. Until that happens, "open" is a promise, not a download.

The training notes say the coding RL ran on hundreds of thousands of long-horizon tasks, and that many tasks used an agentic grader that judged how the model approached the problem, beyond unit tests passing. The limitations section is plain: it is not the most succinct communicator, does not always pick the best tool in codebases with many skills, has no audio, and is less polished as a chat model.

Mistral Studio

Mistral is expanding Studio with managed agents and apps, alongside workflows, enterprise context, agent observability and evaluations. The pitch is one platform where an agent investigates and acts, a workflow coordinates approvals, and an app gives staff a way in, deployable in the cloud, a VPC or on-premises. It builds on Mistral Large 4, which I covered earlier. This is a product announcement with customer quotes, so I treat the production claims as theirs.

NVIDIA DSX OS

The NVIDIA post covers three open source pieces. AI Cluster Runtime defines version-locked configuration recipes. Cluster Readiness Engine runs NCCL communication and training tests against thresholds and bisects failing groups to find suspect nodes. NVSentinel watches GPU, node and network health and can cordon, drain and remediate. The useful idea is that certification repeats after driver upgrades and node swaps, and goodput, not utilization, is the number to watch.

Google's hypertension study

Google Research describes a smartwatch system that uses PPG and accelerometer signals, with no cuff calibration, to detect a persistent hypertensive state over several days. In a prospective study of 196 participants at five US sites against 24-hour ambulatory monitoring, initial screening reached 65.8% sensitivity, 90.0% specificity and 80.6% PPV, against 55.3% sensitivity for office screening.

Open ASR Leaderboard

Hugging Face added two private splits from Liva AI, Egyptian and Nigerian conversational English. Held-out data is the point, since it is harder to tune against.

What I'd watch next

Whether the Inkling weights land on schedule and under the stated license, and what independent evaluations say once people can run them.