🍄 Mario-HTML-v2-4B

V2: less html tag in Cot than v1.

GGUF / llama.cpp build: wesjos/mario-html-v2-4B-GGUF — Q8_0 · Q4_K_M · F16-mmproj. The bundled WebUI below runs on those quants; the quant table is on that card.

---

🖼️ Demo

Every shot is the bundled WebUI (webui/, theme_preset: editorial) with the CoT toggle switched on — webui/config.json ships enable_thinking: false, which hides the reasoning panel entirely, so flip it if you want to see the traces. The panel above each answer is the raw <think> stream; the page below it is the unmodified render.

Screen recording — one session, 67 s

what is 8.2x9.12/2.3? → good for you → 在干嘛?

Persona

mario,how are you

Opinion

how do you think of pizza with pineapple?

Persona, 中文

喜欢吃pizza吗

---

📤 Output Contract

One root element; everything else is an optional child. Class list verified against webui/static/theme.css:

Raw sample (message.content verbatim, v2 zero-system-prompt greedy)

<div class="ai-view">
<h3>为什么天空是蓝色的？</h3>
<div class="summary">
<p class="lead">阳光中的蓝光被大气中的分子"散射"到四面八方，而我们的眼睛对蓝光更敏感，所以我们看到天空是蓝色的。</p>
</div>
<div class="card">
<h4>三句话解释</h4>
<div class="step-list"><div class="step-item">…</div></div>
</div>
</div>

The <think> trace that precedes this sample contains zero HTML tags — pure plain-text planning ("a summary lead, then a three-item stepper..."). Two classes in the sample are content-only: step-list and ai-view-header are not in theme.css, so they render with default block styling — define your own if you want tighter control. (step-item itself is styled.)

---

🎨 WebUI

cd webui
pip install -r requirements.txt
python server.py --port 8791

Open http://localhost:8791. Dual-stream rendering (collapsible <think> + live HTML), clipboard image paste (Ctrl+V), nine theme presets (default · apple · notion · cyber · editorial · cards_rounded · glassmorphism · soft_shadows · compact) plus live custom CSS, and a raw-source / copy-HTML view per message. Defaults in webui/config.json — an OpenAI-compatible endpoint on http://127.0.0.1:18086 serving the alias mario-html-v2-4b. webui/start.sh is the one-command path (llama.cpp GGUF backend + WebUI, via systemd-run --user); it expects a sibling start_backend.sh, which is not part of this repo.

---

⚠️ Notes

• The output is HTML, not Markdown. Feed it a Markdown renderer and you get raw tags.
• The WebUI ships a non-empty default system_prompt — clear it for the zero-system-prompt path.
• v2's webui/config.json also ships enable_thinking: false, i.e. direct-HTML mode. Turn CoT on to get the traces shown above; the purity claim is about what the traces contain, not about them being on by default.
• The model picks from the class vocabulary above; arbitrary new CSS classes are not a contract. Custom theme.css is the supported way to restyle.
• Arithmetic is reliable but not bank-grade — check long decimal chains.
• --reasoning-parser qwen3 needs a recent vLLM.
• webui/run_webui.sh is Linux-oriented (hardcoded /mnt/sdb/... interpreter and TMPDIR paths). config.json itself is portable (./model.q8_0.gguf); adjust the script for another OS.

---

📄 License

Apache License 2.0.