| --- |
| title: Vontra Labs |
| sdk: static |
| pinned: false |
| --- |
| |
| <div align="center"> |
|
|
| # Vontra Labs |
|
|
| **PhaseMesh: compact local AI runtimes built from measured traces, phase signatures, and hard evals.** |
|
|
| [PhaseMesh-Qwen3-4B](https://huggingface.co/Vontra/PhaseMesh-Qwen3-4B) · [Source](https://github.com/ashhart/PhaseMesh) |
|
|
| </div> |
|
|
| --- |
|
|
| ## PhaseMesh |
|
|
| PhaseMesh is a compact runtime for prompt-conditioned coding and reasoning traces. |
|
|
| The current public artifact pours Qwen3-4B-Instruct traces into a local PhaseMesh memory, then serves covered skills through phase-signature retrieval with an explicit confidence gate. It is built to be fast, inspectable, and honest about scope. |
|
|
| ## Current Artifact |
|
|
| | Metric | Result | |
| | --- | ---: | |
| | Artifact | [PhaseMesh-Qwen3-4B](https://huggingface.co/Vontra/PhaseMesh-Qwen3-4B) | |
| | Teacher traces | 180 | |
| | Held-out coverage eval | 65/65 | |
| | Median PhaseMesh latency | ~2.18 ms | |
| | Live Qwen3-4B CUDA answer latency | ~4,966 ms median | |
|
|
| The result is a fast specialist over a published coverage set. It is **not** a raw Qwen checkpoint, and it is **not** claimed as general LLM parity. |
|
|
| ## What We Care About |
|
|
| - Small local AI systems that can be inspected and shipped. |
| - Trace-poured specialists that answer covered workflows in milliseconds. |
| - Evaluation files and artifacts that can be audited, not just screenshots. |
| - Architecture work on phase binding, associative recall, and efficient sequence kernels. |
|
|
| ## Links |
|
|
| - Model: [Vontra/PhaseMesh-Qwen3-4B](https://huggingface.co/Vontra/PhaseMesh-Qwen3-4B) |
| - Source: [github.com/ashhart/PhaseMesh](https://github.com/ashhart/PhaseMesh) |
|
|
