MakoBench Suite History: Earlier Versions
- Category
- AI & Local LLM
- Posted
- November 5, 2026
- By
- Jacob Lloyd — written with AI assistance, post-project
- Read time
- 1 min read
In plain terms: MakoBench scores only mean something against the version of the tests that produced them. When the tests change, we re-run the models and keep the old board exactly as it was. This page lists each suite version and where to find its board.
Go to: CrucibleForge on GitHub
The tests behind MakoBench change from time to time. A score only means something against the suite that produced it, so we never compare numbers across suites and we never quietly rewrite an old board.
Why do suites change?
We drop tests that turn out to be slow or noisy, fix questions that were unfair, and add harder sets as models improve. When a change moves scores, every model is re-run on the new suite. The previous board is then kept exactly as it stood, with its own margins, its own scale and the date it was built.
Where are the earlier boards?
The Main Bench, +NSFW Bench and The tests pages each have a Suite version menu at the top. Pick an earlier version there and the page shows that suite as it stood.
Versions
- Suite 4.0 (current). Total is Code plus Chat, 10 points each on the standard tests. See the Main Bench.
- Suite 3.0. Scored out of 100 on the old tests, so do not set these numbers beside suite 4.0. The archived board has 32 models and was built from the board we archived on 2026-09-23. Pages: Main Bench, +NSFW Bench and The tests.
What changed between them?
The write-up for the current suite is Suite 4.0: what changed. For every release of the engine, including ones that did not touch the scores, read the CrucibleForge changelog on GitHub.
Related: How MakoBench works.