AI Technology Observations | 4 November 2026, 07:00

This observation covers verified technical changes made public from approximately 7:20 am on 3 November to 7:20 am on 4 November 2026. It includes two items.

Sarvam Prism ships with open weights and a Decisions API

Sarvam AI released Sarvam Prism on 3 November. Built on Sarvam 105B, it is a decision model that takes a state, a question and a fixed set of options, then returns a probability for every option in one forward pass instead of generating free-form text. Compared with the earlier general-purpose Sarvam 105B release, the new delivery is a specialised interface for classification, routing, moderation, agent-run judging and security checks. The weights are available through Hugging Face under the Apache 2.0 licence, while the hosted version is served through the Decisions API at POST /v2/systemone.

The hosted API accepts text or JSON and does not support images or streaming. State, question and options are limited to 32,768 tokens per question, and a choice question can contain 2 to 255 options. The open weights use FP8; Sarvam says self-hosting requires an FP8 GPU with about 120 GB of free memory. Sarvam reports an average score of 71.8 across three decision indices and a median response time of 44 milliseconds on one B200. These are developer evaluations: one index is Sarvam's recreation, and the Indian-language and cybersecurity results include internal data. There was no public independent reproduction by the cut-off, and Sarvam says the cybersecurity evaluation set will be released later.

Google and Mercor open the PLE engineering-agent benchmark and competition

Google AI & Infrastructure and Mercor opened 29 public development tasks, grading tools, an evaluation container and the agent competition for Physical-Engineering's Last Exam (PLE) on 3 November. The PLE paper was first submitted on 21 October and revised on 30 October. The new fact in this window is the public delivery of tasks and tools and the opening of submissions, rather than the paper's initial appearance.

PLE contains 85 long-horizon tasks across mechanical, electrical, civil and structural engineering, with 29 public development tasks and 56 hidden test tasks. Each task runs in an isolated container with project files that may include PDFs, spreadsheets, email, 2D CAD drawings, 3D BIM models and revision records. Entrants must produce deliverables such as calculation packages, design reviews, drawings or technical memoranda. The competition runs each submission three times on the hidden tasks and ranks the average. Google's announcement reports a best native-harness baseline of 48.3%, while the current competition page separately lists an OpenHands configuration at 48.6%. Both are organiser-run baselines using different agent configurations and do not establish a model-wide capability. The full competition timetable and prize details remain unannounced, and the hidden tasks are not public.

Sources

Sarvam AI, Introducing Sarvam Prism, 3 November 2026, https://www.sarvam.ai/blogs/sarvam-prism

Sarvam AI Documentation, Sarvam Prism, accessed 3 November 2026, https://docs.sarvam.ai/api/token-factory/models/sarvam-prism

Hugging Face, sarvamai/prism, published 3 November 2026, https://huggingface.co/sarvamai/prism

Google Cloud Blog, Introducing Physical-Engineering's Last Exam for agents, 3 November 2026, https://cloud.google.com/blog/products/ai-machine-learning/introducing-physical-engineerings-last-exam-for-agents

Google Research, Physical-Engineering's Last Exam repository, accessed 3 November 2026, https://github.com/google-research/ple

Physical-Engineering's Last Exam, Competition, accessed 3 November 2026, https://ple-bench.com/competition