Laser and other tech projects,
always free for your use.
Every post has the details, process, and exact settings to duplicate it yourself. The site is live and maintained — and most of the work here, including this site itself, is done with LLMs.
MakoBench
Language models scored on code and chat with one open test suite, run on our own hardware. Suite 4.1.
- Kimi K3.1 Pro 24.6
- Claude Fable 5.2 24.2
- GPT-6 Astra 24.0
Latest projects
Builds, guides, and experiments from the shop.
Claude Fable 5.2: 24.2 on MakoBench, tied for first
Claude Fable 5.2 scores 24.2 on MakoBench 4.1, tied for first with Kimi K3.1 Pro, GPT-6 Astra and GPT-6.1 Sol, and up from 19.4 for Fable 5.1.
MakoBench Suite History: Earlier Versions
The MakoBench tests change from time to time, and a score only means something inside the suite that produced it. Here is where the earlier boards live and why we kept them.
MakoBench 4.1: What Changed
Suite 4.1 is a correction release with no new cases. Two independent model reviewers read every public answer key, about 190 problems were confirmed, and the board has been re-run: 31 of the 38 models scored on both boards rose, on both Chat and Code.
MakoBench Suite 4.0: Nominal 20, Class and Lean
Suite 4.0 scores Code and Chat at 10 each instead of a percentage, drops the slowest and noisiest tests, and adds Class and Lean. Here is what the public release changed and what is in our own copy of the tool but not yet in a release.
My AI Agent Built a Benchmarking Tool for My AI Agents
Our coding agent wrote bench-llm: 1,660 lines of Python that benchmark local models in LM Studio. Within two weeks our models all scored 5/5, so we replaced it with CrucibleForge: 276 cases, 166 hard, free on GitHub.
How MakoBench Works: Open LLM Benchmark Explained
How MakoBench turns a model's answers into a score, step by step, with a diagram for each: how cases are graded, how the judge works, Code plus Chat, the margin, and how bonus sets unlock. Worked examples use real models from the board.
MakoBench Adds a Run Cost Column
Hosted models on the MakoBench board now show what one run of the standard Code and Chat tests cost. Replies a provider's content filter blocks are marked on the row and score zero.
How CrucibleForge Works: An Open LLM Benchmark, Explained
CrucibleForge is the free, open-source benchmark behind the MakoBench board. Five blocks, one diagram: your model, a setup check, the tests, the graders and the scores. Plus a ten-minute guide to running it on a local server or a hosted API.
What MakoBench Tests, and What It Doesn't
MakoBench tests text models on two jobs, working code and role-play or companion chat. Here is which models belong on the board, which never will, and what a row needs before it gets a rank.
Browse by type
AI & Local LLM
Running AI on your own hardware — local LLMs, agents, and image generation. No subscriptions, no cloud, no computer-science degree.
Laser Projects
Cutting, engraving, and the occasional small fire. Each post has the details, process, and settings you need to duplicate it yourself.
Tools & Downloads
Small apps and utilities we built and actually use. Free downloads for personal use, source included.