Product boundary (read this first)
- Middleware Pro V5 — commercial integration/productivity product for IDEs (VS Code, Cursor, etc.).
- llama.cpp TRA — modified experimental runtime that implements TRA learn paths.
- Ministral 14B TRA — model/research package prepared to demonstrate TRA with that runtime.
Middleware Pro does not confer /learn, forward learning, or /sleep on every local GGUF. Those capabilities require the TRA runtime and a compatible TRA model package.
What TRA learning modifies
TRA learning in the current experimental stack operates through the TRA runtime’s LoRA / learn adapter path, not by silently rewriting an arbitrary third-party GGUF as a general property of Middleware Pro.
In practice, focused learning sessions produce persistent TRA learn artifacts associated with that runtime session and model — notably adapter and optimizer snapshot files (for example adapter.bin / optimizer.bin with metadata), which can be reloaded in later TRA sessions.
| Mechanism | What changes | What does not automatically change |
|---|---|---|
/learn (focused) |
TRA learn/adapter state for the supplied material, under the TRA runtime | Your IDE; unrelated GGUFs loaded in stock llama.cpp/Ollama without TRA |
| Forward learning | Learn activity may continue while interaction continues (TRA runtime) | Middleware Pro configuration alone |
/sleep |
Consolidation of previously acquired TRA learn state into the persistent TRA learn/adapter state used later by that stack | A blanket claim that every base GGUF file on disk is rewritten for all users/tools |
Base weights / GGUF consolidation
Claims that learning has been consolidated into base model weights or a redistributable GGUF are substantial. We do not use marketing language that implies every TRA session rewrites the published Ministral base GGUF for all tools.
Where a workflow exports or merges learn state into a durable artifact, that artifact type (adapter snapshot vs merged weights vs other derived state) will be stated explicitly in the release notes for that package, with enough detail for technical reproduction. Until a specific release note says otherwise, treat persistence as TRA learn/adapter state under the TRA runtime.
What we will publish with the research packages
- Clear package labels: runtime vs model vs Middleware Pro.
- Commands and env flags required for TRA learn on the supported stack.
- Description of on-disk artifacts produced by
/learn//sleepfor that release. - No implication that Middleware Pro turns every GGUF into a TRA learner.
Status of public downloads: available — llama.cpp TRA, Middleware Pro V5 (eval), Ministral weights via Hugging Face.
TRA Research Paper — Preprint, Version 1.0 · Back to products · Downloads