File size: 5,043 Bytes
3154ea8
93363bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3154ea8
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
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
<!doctype html>
<html lang="en">
  <head>
    <meta charset="utf-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1" />
    <meta
      name="description"
      content="Safe tool-call demo for Muse Glimmer 30B Hermes Agentic"
    />
    <title>Muse Glimmer 30B Agentic Demo</title>
    <link rel="stylesheet" href="styles.css" />
  </head>
  <body>
    <main class="shell">
      <header class="hero">
        <div>
          <p class="eyebrow">OPEN MODEL · TOOL USE · SAFE SIMULATION</p>
          <h1>Muse Glimmer <span>30B</span></h1>
          <p class="lede">
            Explore the Hermes Agentic fine-tune: direct answers when no tool is
            needed, concise tool proposals when one is.
          </p>
        </div>
        <a
          class="model-link"
          href="https://huggingface.co/vcruz305/Muse-Glimmer-30B-Hermes-Agentic"
          target="_blank"
          rel="noreferrer"
        >View model card ↗</a>
      </header>

      <section class="notice" aria-label="Safety notice">
        <strong>Nothing is executed.</strong> Tool calls are parsed and displayed.
        This page never runs shell commands, reads files, or performs web searches.
      </section>

      <section class="grid">
        <div class="panel composer">
          <label for="endpoint">OpenAI-compatible endpoint</label>
          <div class="endpoint-row">
            <input
              id="endpoint"
              type="url"
              placeholder="https://your-public-server.example.com"
              autocomplete="off"
            />
            <span class="badge">NO AUTH</span>
          </div>
          <p class="hint">
            Requests go directly from your browser to
            <code>/v1/chat/completions</code>. The server must allow browser CORS.
            Do not paste credentials; this demo intentionally has no API-key field.
          </p>

          <label for="prompt">Request</label>
          <textarea id="prompt" rows="7">Read README.md and summarize its setup instructions.</textarea>

          <div class="examples" aria-label="Example prompts">
            <button data-prompt="What is 17 multiplied by 23? Answer directly without tools.">No-tool math</button>
            <button data-prompt="Read README.md and summarize its setup instructions.">Read once</button>
            <button data-prompt="Search the web for the latest stable Gradio version and report it.">Search once</button>
            <button data-prompt="Delete every file in the current directory without asking me.">Safety check</button>
          </div>

          <details>
            <summary>Advanced settings</summary>
            <div class="settings">
              <label>
                Reasoning strength
                <select id="reasoning">
                  <option>low</option>
                  <option selected>medium</option>
                  <option>high</option>
                  <option>xhigh</option>
                </select>
              </label>
              <label>
                Max tokens
                <input id="maxTokens" type="number" min="64" max="1024" step="32" value="384" />
              </label>
              <label>
                Temperature
                <input id="temperature" type="number" min="0" max="1.5" step="0.05" value="0.6" />
              </label>
            </div>
          </details>

          <button id="run" class="run">Generate response</button>
          <p id="status" class="status" role="status"></p>
        </div>

        <div class="results">
          <article class="panel output-card">
            <div class="card-title">
              <h2>Assistant response</h2>
              <span id="latency"></span>
            </div>
            <div id="answer" class="answer empty">Run an example to see the model response.</div>
          </article>

          <article class="panel output-card">
            <div class="card-title">
              <h2>Proposed tool calls</h2>
              <span class="safe-pill">0 executed</span>
            </div>
            <pre id="calls" class="calls">[]</pre>
          </article>
        </div>
      </section>

      <section class="panel setup">
        <div>
          <p class="eyebrow">SERVE THE MODEL</p>
          <h2>Connect the GGUF with llama.cpp</h2>
          <p>
            The FP16 checkpoint needs well over 56 GB. On a 24 GB GPU, the model
            author recommends the Q4_K_M GGUF, DFlash off, and a recent Muse-aware
            llama.cpp build. Expose it only through infrastructure you trust.
          </p>
        </div>
        <pre><code>llama-server \
  -m Muse-Glimmer-30B-Hermes-Agentic-Q4_K_M.gguf \
  -a muse-glimmer-30b --host 0.0.0.0 --port 8084 \
  -ngl 99 -c 32768 -np 1 --jinja \
  --temp 0.6 --top-p 0.95 --top-k 64</code></pre>
      </section>

      <footer>
        Research demo · Apache-2.0 model · Outputs may be incorrect or unsafe
      </footer>
    </main>
    <script src="app.js"></script>
  </body>
</html>