Question: where would a flash-resident tiny language runtime fit relative to TNN’s unified inference model?

[Alpha-Guardian]

Hi TNN folks,

I wanted to share a small edge-language-runtime experiment and ask how people here would think about it relative to the more familiar unified graph/runtime view of on-device inference.

We built a public demo line called Engram and deployed it on a commodity ESP32-C3.

Current public numbers:

• Host-side benchmark capability

  • LogiQA = 0.392523
  • IFEval = 0.780037

• Published board proof

  • LogiQA 642 = 249 / 642 = 0.3878504672897196
  • host_full_match = 642 / 642
  • runtime artifact size = 1,380,771 bytes

Important scope note:

This is not presented as unrestricted open-input native LLM generation on MCU.

The board-side path is closer to a flash-resident, table-driven runtime with:

• packed token weights
• hashed lookup structures
• fixed compiled probe batches
• streaming fold / checksum style execution over precompiled structures

So this is not a standard dense graph flowing through a conventional unified inference stack. It is closer to a task-specialized language runtime whose behavior has been pushed into a very constrained executable form.

Repo:
https://github.com/Alpha-Guardian/Engram

What I’m curious about is whether systems like this should be thought of as:

• outside the normal unified-inference framework scope
• an extreme endpoint of edge specialization
• or an adjacent class of language-task systems that future deployment frameworks may need to account for

Would be interested in any thoughts.