ashxhart commited on
Commit
17f37c1
路
verified 路
1 Parent(s): 59c9b17

Update organization card

Browse files
Files changed (1) hide show
  1. README.md +10 -41
README.md CHANGED
@@ -1,47 +1,16 @@
1
- ---
2
- title: Vontra Labs
3
- sdk: static
4
- pinned: false
5
- ---
6
 
7
- <div align="center">
8
 
9
- # Vontra Labs
10
 
11
- **PhaseMesh: compact local AI runtimes built from measured traces, phase signatures, and hard evals.**
12
 
13
- [PhaseMesh-Qwen3-4B](https://huggingface.co/Vontra/PhaseMesh-Qwen3-4B) 路 [Source](https://github.com/ashhart/PhaseMesh)
14
 
15
- </div>
 
 
 
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.