Muse Glimmer's weak LiveCodeBench score traced to odd Numba dependency

dejavucoder · x · 2026-08-20

Evaluator @bnjmnmarie found Muse Glimmer scored only 76.2% (804/1,055) on LiveCodeBench, about 14 points behind Gemma 4. Digging in, 175 failures came from the model's habit of using Numba—a library EvalScope's official sandbox doesn't include. Across 100+ models evaluated, he had never seen one rely on Numba like this, and the benchmark problems don't need it.

Installing Numba wasn't enough: compilation blew up memory and timed out even with a 120-second limit (vs. his usual 10), requiring a different machine. Accuracy recovered to 90% once fixed. A reminder that a model's real capability can be masked by evaluation-environment assumptions.

Original post →

More from Models

Models channel →