Ulam ranks Mistral Large 4 seventh on STEMBench

Ulam’s STEMBench120 page labels its audit snapshot Oct. 28, 2026, and lists Mistral Large 4 with a result of 2.501 and a rank of seven.[2] Ulam identifies engineering and astronomy as the model’s strongest domains in the run.[2] The benchmark lists biology, chemistry, physics, astronomy, Earth science and engineering as its six domains.[2] Its full run covers 120 research tasks, with 20 tasks in each domain.[3]

Ulam says STEMBench’s fractional scores describe progress, not a percentage of tasks solved or a probability that an answer is correct.[2] The 2.501 result should therefore be read as a benchmark score, not as a claim that the model solved a corresponding share of tasks.[2] Ulam describes the work as research tasks, and the result is tied to a run covering 120 tasks.[3][2] The snapshot does not, in the supplied information, provide task-level judgments or submissions for public comparison.[2]

Ulam lists Mistral Large 4’s STEMBench run as closed-book.[2] Of its 120 scored attempts, 118 contained solver-authored content and two repeated the input.[2] Ulam reports 111 level-2 results and seven level-1 results, with no level-3 or full-target results for the model.[2] These counts describe the outcomes of this run; they do not establish that the model completed a full research target.[2]

Ulam reports that none of the 11 runs in the Oct. 28 snapshot met a full task target.[2] It also keeps STEMBench120 task statements, per-task judgments and model submissions out of the public comparison.[2] That means readers can compare the listed result and rank, but the supplied public comparison does not expose those underlying materials for independent task-by-task review.[2] The result is evidence about performance on this benchmark run, not a complete account of the model’s capabilities.[3][2]