Spaces:
Running
Running
Update organization card
Browse files
README.md
CHANGED
|
@@ -1,47 +1,16 @@
|
|
| 1 |
-
|
| 2 |
-
title: Vontra Labs
|
| 3 |
-
sdk: static
|
| 4 |
-
pinned: false
|
| 5 |
-
---
|
| 6 |
|
| 7 |
-
|
| 8 |
|
| 9 |
-
|
| 10 |
|
| 11 |
-
|
| 12 |
|
| 13 |
-
|
| 14 |
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
-
-
|
| 18 |
-
|
| 19 |
-
## PhaseMesh
|
| 20 |
-
|
| 21 |
-
PhaseMesh is a compact runtime for prompt-conditioned coding and reasoning traces.
|
| 22 |
-
|
| 23 |
-
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.
|
| 24 |
-
|
| 25 |
-
## Current Artifact
|
| 26 |
-
|
| 27 |
-
| Metric | Result |
|
| 28 |
-
| --- | ---: |
|
| 29 |
-
| Artifact | [PhaseMesh-Qwen3-4B](https://huggingface.co/Vontra/PhaseMesh-Qwen3-4B) |
|
| 30 |
-
| Teacher traces | 180 |
|
| 31 |
-
| Held-out coverage eval | 65/65 |
|
| 32 |
-
| Median PhaseMesh latency | ~2.18 ms |
|
| 33 |
-
| Live Qwen3-4B CUDA answer latency | ~4,966 ms median |
|
| 34 |
-
|
| 35 |
-
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.
|
| 36 |
-
|
| 37 |
-
## What We Care About
|
| 38 |
-
|
| 39 |
-
- Small local AI systems that can be inspected and shipped.
|
| 40 |
-
- Trace-poured specialists that answer covered workflows in milliseconds.
|
| 41 |
-
- Evaluation files and artifacts that can be audited, not just screenshots.
|
| 42 |
-
- Architecture work on phase binding, associative recall, and efficient sequence kernels.
|
| 43 |
-
|
| 44 |
-
## Links
|
| 45 |
-
|
| 46 |
-
- Model: [Vontra/PhaseMesh-Qwen3-4B](https://huggingface.co/Vontra/PhaseMesh-Qwen3-4B)
|
| 47 |
-
- Source: [github.com/ashhart/PhaseMesh](https://github.com/ashhart/PhaseMesh)
|
|
|
|
| 1 |
+
# Vontra
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
+
Vontra is a small, one-person AI tinkering space focused on practical local inference.
|
| 4 |
|
| 5 |
+
I use this space to share experiments that may be useful to other people running models on their own hardware: MLX conversions, quantized checkpoints, serving notes, and lightweight tools.
|
| 6 |
|
| 7 |
+
The work here is experimental by nature. Models and notes are shared in the open so others can test, compare, remix, and learn from them.
|
| 8 |
|
| 9 |
+
## What You Will Find Here
|
| 10 |
|
| 11 |
+
- Local AI experiments
|
| 12 |
+
- Model conversions and quantizations
|
| 13 |
+
- Apple Silicon and MLX testing
|
| 14 |
+
- Inference notes from real hardware
|
| 15 |
|
| 16 |
+
Everything is published best-effort, with attribution back to the original model authors where applicable.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|