Skip to content

Lab · Industrial model evaluations

Industrial model evaluations

Dated records of how open-weight models behave on tasks drawn from maintenance, inspection, alarms and routing. Each record states who ran it, what was checked, what the counts support and what they leave open. None is a leaderboard.

Back to current Lab work

Records

  • AI-run desk study · 19 November 2026

    When a valid alarm label misses the alarm

    Olala v1.0 mini, Ministral 3 3B and 8B, and Qwen3.5-9B answered sensor and fault questions and simulated plant tasks from 8-bit GGUF files on a cloud CPU guest using four threads. Three also classified inspection photographs. The record identifies the separate hosted trials.

    Read the record

How to read a record

Status label
Each record says who ran the evaluation and on what machine. A desk study run by an AI assistant is labelled as one.
Counts, not percentages
Results are correct answers over attempts, with the trivial baseline beside them where one exists.
Serving profile
A result belongs to a model, a file, a runtime and an output allowance together. Change one and the result can change.
Trial, not winner
A record names candidates worth testing on your cases. It does not predict performance on your plant.

In preparation

A qualification record run on Hyperion's own workstation is in preparation as a separate record.