godqhr1421's picture
Upload index.html
93363bc verified
Raw
History Blame Contribute Delete
5.04 kB
<!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>