File size: 4,665 Bytes
5322285
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
<!doctype html>
<html lang="en">
	<head>
		<meta charset="utf-8" />
		<meta name="viewport" content="width=device-width, initial-scale=1" />
		<title>SAURON — a model trained by its own fleet · TibiOS</title>
		<meta name="description" content="SAURON: a Qwen3-4B LoRA adapter trained by the TibiOS fleet itself — five home machines, zero cloud, zero token spend." />
		<link rel="icon" href="assets/tibios-logo.png" />
		<link rel="stylesheet" href="style.css" />
	</head>
	<body>
		<header class="nav">
			<div class="brand">
				<img src="assets/tibios-logo.png" alt="TibiOS logo" class="logo" />
				<span>TibiOS</span>
			</div>
			<nav>
				<a href="index.html#how">How it works</a>
				<a href="index.html#fleet">The fleet</a>
				<a href="index.html#demo">Demo</a>
				<a href="sauron.html">Sauron</a>
				<a href="index.html#contact">Contact</a>
			</nav>
			<a class="btn btn-small" href="https://tibios.ai">tibios.ai</a>
		</header>

		<section class="hero">
			<p class="tagline">Presenting</p>
			<h1>SAURON</h1>
			<p class="tagline">A model trained by its own fleet.</p>
			<p class="lede">
				What if a set of home computers could generate their own training
				data, train a model, deploy it and use it for inference — all within
				the same infrastructure? That is what we achieved with TibiOS.
				<strong>Zero token spend.</strong>
			</p>
		</section>

		<section class="section">
			<h2>The fleet that trained it</h2>
			<p class="lede center">
				SAURON is based on Qwen3-4B, adapted with LoRA on a heterogeneous
				fleet of five machines:
			</p>
			<div class="cards">
				<div class="card"><h3>🍎 Mac mini M4</h3><p>16 GB — training with MLX.</p></div>
				<div class="card"><h3>💻 MacBook Air M1</h3><p>16 GB — training with MLX.</p></div>
				<div class="card"><h3>⚡ 2 × Jetson Orin Nano</h3><p>8 GB each — training with Unsloth + TRL on CUDA.</p></div>
				<div class="card"><h3>🔌 Orange Pi 6+</h3><p>16 GB — head node and router.</p></div>
			</div>
		</section>

		<section class="section">
			<h2>The full pipeline</h2>
			<div class="steps">
				<div class="step"><span>1</span><p>The inference lanes generate their own training data.</p></div>
				<div class="step"><span>2</span><p>The Macs train in a ring, splitting each batch by every machine's measured speed.</p></div>
				<div class="step"><span>3</span><p>Training produces the SAURON LoRA adapter.</p></div>
				<div class="step"><span>4</span><p>The head node publishes the adapter as a new inference lane: <strong>@tuned</strong>.</p></div>
				<div class="step"><span>5</span><p>The fleet infers again — and can generate new data for the next cycle.</p></div>
			</div>
		</section>

		<section class="section">
			<h2>Technical data</h2>
			<div class="cards">
				<div class="card"><h3>Qwen3-4B · 4-bit</h3><p>Base model, quantized.</p></div>
				<div class="card"><h3>MLX + Unsloth/TRL</h3><p>MLX on Apple Silicon; Unsloth + TRL on CUDA.</p></div>
				<div class="card"><h3>LoRA r=8, α=16</h3><p>Rank 8, alpha 16, seven projections.</p></div>
				<div class="card"><h3>16.5M parameters</h3><p>Trainable — about 0.74% of the model.</p></div>
				<div class="card"><h3>64 MB</h3><p>Adapter size.</p></div>
				<div class="card"><h3>Home Wi-Fi LAN</h3><p>Zero electricity cost in our scenario, no cloud needed for this local pipeline.</p></div>
			</div>
			<p class="lede center note">
				Kimi K3 and Claude were used exclusively to draft the glossary data —
				not to train SAURON.
			</p>
		</section>

		<section class="section">
			<img class="wide" src="assets/sauron-fleet-training.png" alt="SAURON infographic: the fleet, the pipeline and the LoRA adapter data" />
		</section>

		<section class="section">
			<p class="lede center big-quote">
				The important result is not just a tuned model. It is a working cycle
				in which the same fleet both trains and infers — a fundamental piece
				of the TibiOS vision: distributed AI infrastructure where heterogeneous
				machines collaborate, learn and serve models.
			</p>
			<p class="lede center big-quote">
				<strong>You don't need to start with a data center.<br />Start with what you have.</strong>
			</p>
		</section>

		<footer id="contact" class="footer">
			<img src="assets/tibios-logo.png" alt="" class="logo" />
			<p><strong>TibiOS</strong> — Decentralized AI for everyone</p>
			<p>
				<a href="https://tibios.ai">https://tibios.ai</a> ·
				<a href="mailto:fgparamio@tibios.ai">fgparamio@tibios.ai</a> ·
				+52 999 122 3748
			</p>
			<p class="tiny">Global community — we collaborate, we grow, we innovate.</p>
		</footer>
	</body>
</html>