TatuEngine — the engine that never stops
TatuEngine·

TatuEngine — the engine that never stops

3 min read← Back to timeline
#tatuengine#ssm#maturity#stability#engine

Some projects live off commit explosions — new features every day, monumental refactors, roadmaps that change every sprint. The TatuEngine is not one of them.

In the last 30 days, the repository accumulated 50 commits. 97% are auto-sync. Zero new features. Zero fixes. The engine simply works.

And that, I believe, is the greatest indicator of maturity a project can have.

What 298 commits built

The TatuEngine started as an experiment: “what if I train an SSM from scratch, in my own format, with GPU inference?” — the question that had been following me since I read the Mamba paper.

298 commits later, the project is a full engine with:

  • GPU inference with BitMamba-2 1B — block-encoded BVH with radical compression of ternary models
  • GPU block-traversal kernel — parallel traversal with hundreds-of-times speedup over CPU
  • Hybrid CPU/GPU pipeline — SSM state on CPU, bridge_matmul on GPU
  • tatu CLI — generation, MCP and tool commands
  • MCP connector — standalone server with 4 tools (FastMCP, stdio/SSE)
  • Autopoietic agent — GenerationContext with 8 subsystems

The C++23 + CUDA engine is constantly evolving. C++ tests are being restructured — the current focus is more on stability than new features.

The calm after the storm

What changed in the last 30 days? Almost nothing visible. What changed was the background:

  • Weekly docs auto-sync: refinement, not rewriting
  • Automatic backups: functional checkpoint in bare repo
  • No critical bugs reported since July

The project has entered what I call cruise phase — the engine doesn’t need new features because it already does what it promises. The questions now are of a different nature:

  • How to integrate TatuEngine into Arachne as an inference backend?
  • Does BitMamba-2 1B need fine-tuning for specific tasks?
  • Does the security sandbox need reimplementation in the new test architecture?

What’s next

Cruise phase isn’t stagnation. It’s time to build on top of, not inside. TatuEngine as an inference backend for the Arachne ecosystem is the next logical step.

The engine is ready. The journey continues.