Mario-HTML-v2-9B is a 9.65B-parameter fine-tune of Qwen3.5-9B (bf16, about 18 GiB, 262,144-token native context) that answers in HTML rather than Markdown, taking text or text plus an image and producing a styled page of cards, grids, metric tiles, steppers, callouts, badges, tables and code blocks. It is tuned to work with no system prompt, which the card calls the trained path, and its output follows a fixed contract: one <div class="ai-view"> root and a fixed vocabulary of CSS classes. Restyling is done by supplying your own theme.css, not by expecting new classes. The v2 change is that the <think> reasoning trace contains fewer HTML tags than v1, and the card's examples show it as plain-text planning with none. The card shows demos of a microservices-versus-monolith comparison grid, a step-by-step linear-equation solution and a llama photo answered with a llama-versus-alpaca breakdown. A WebUI that ships with the sibling mario-html-v2-4B repo renders the reasoning stream and the live HTML, with nine theme presets, clipboard image paste and a copy-HTML view. There are no benchmark numbers. The notes warn that arithmetic is reliable but not bank-grade, so long decimal chains need checking, and that --reasoning-parser qwen3 needs a recent vLLM. A GGUF build for llama.cpp (Q8_0 and Q4_K_M, plus a BF16 mmproj) is available separately, and the model is released under Apache 2.0.
MTP — Speculative Decoding: Yes
Multi-Token Prediction (MTP) GGUF is a specialized GGUF model file format extension that integrates speculative decoding directly into the model weights to significantly accelerate local inference. Unlike traditional speculative decoding which requires a separate, smaller "draft" model, MTP GGUF files include additional output heads within the main model architecture that predict multiple future tokens in a single forward pass.
