Crucible
Crucible is an AI task gauntlet to analyze whether local models are capable of real work. Can they write code for personal projects? Keep working when the internet is unavailable? Accomplish tasks involving personal data we don't want to share with cloud providers for training? The tasks are divided into chat and agentic sections and are categorized into three tiers of difficulty. Play around, I think you'll be surprised by the performance of these small models.
Everything below was created and written by or with the help of AI.
Crucible runs the same real-world prompts in 2 separate sections that are never ranked together. Chat tests (writing, math & logic, geography, history, general knowledge, evals): One API call, no tools. Agentic tests (coding, research, everyday): An agent harness with tools and a workspace, harness: opencode or pi. Each run is scored 0–10 against the test's rubric. Tests come in tiers: Spark (the basics, pass or fail), Temper (everyday work, the headline score), Forge (frontier-level work). How Crucible works, and how to run it yourself →
Loading results…
Leaderboard
Quality vs time
Every model on every test
Score per test, grouped by tier. Rows keep the leaderboard's order. Cells with a green corner are scored entirely by programmatic checks (no judge); select any cell for its checks or the judge's note. Dashed cells are provisional tests that don't count toward tier scores yet.
Model × area
Mean score (0–10) for each area in this section, across all tiers. Where a model has run none of an area's active tests yet, a dashed cell shows its provisional tests instead.
One model in detail
Latest run per test, grouped by area. Each collapsed area shows its mean score, how many of its tests ran, and one square per test, easiest tier first. Open an area for its tests, then select a test for its checks or the judge's summary.