The industry benchmark just started measuring agents, not just models. mlperf inference v6.1, out…
the industry benchmark just started measuring agents, not just models. mlperf inference v6.1, out september 16, adds two new tests for agentic inference on top of a record participation count.
the unit of compute being benchmarked is now the run, not the prompt.
Context
MLCommons' release of 16 September 2026 says MLPerf Inference v6.1 introduces two new tests aligned with recent deployment trends including agentic inference: End-to-End Retrieval-Augmented Generation, a multi-model pipeline with an embedding model, retriever, re-ranker and LLMs, and Edge Agentic Inference, covering multi-turn workloads such as agentic coding with growing conversational history. It adds speculative decoding support in the interactive scenario, reports a record 30 participating organizations, and reports up to 5.7 times performance gain over one year earlier on DeepSeek R1 per-accelerator in the server scenario and 2.99 times on the VLM test versus v6.0.
The date, the two new tests and the 30 organizations are first-party. Two new tests for agentic inference is overstated: only Edge Agentic Inference is explicitly agentic, and the other is an end-to-end RAG pipeline that MLCommons describes as multi-step and multi-component. MLPerf measures serving performance of hardware and software on fixed workloads and not an agent's task success. That the unit being benchmarked is now the run and not the prompt is the author's gloss and not a statement in the release.
Watch next
- The datacenter-side agentic workload and the published Edge Agentic scenario details.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 18:44 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →