---
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)