---

license: apache-2.0

base_model:

- Logics-MLLM/Logics-PPT-Qwen3.6-35B-A3B-RL

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

language:

- zh

- en

tags:

- presentation-generation

- slides

- html-generation

- gguf

- llama.cpp

- 42model

---

<!-- README for the Hugging Face repo 42ailab/Logics-PPT-Qwen3.6-35B-A3B-RL-GGUF. Editing this file changes the repo landing page. -->

Logics-PPT · Turn a Block of Content into One HTML Slide

English | 简体中文 · Also on ModelScope · 42ailab.com

Paste the content for one slide (a title, some points, a few figures) and it produces a single 16:9 HTML slide you can open in a browser. Everything runs on your own machine.

> The model was built by the Alibaba Logics team and released under Apache-2.0 (Logics-PPT-Qwen3.6-35B-A3B-RL, a 35B-total / ~3B-active MoE). This repository is not a new model: it is a quantized GGUF we converted from the official weights, plus our own test results. The capability belongs upstream.

What it is for

When you prepare a presentation, the text is usually written already; the time goes into layout: colours, alignment, cards, making the numbers stand out. That is the step this model was trained for. Upstream refined it with reinforcement learning, scoring rendered slides on five layout measures (aspect ratio, whitespace, no overlapping elements, visual balance, readable font sizes).

It makes one slide at a time. The intended use is: you have one slide's worth of content and want a decent layout to start from.

What we saw in testing

15 runs of Q4_K_M on an Apple M3 Max (the full log is in our repository):

| Case | Result |

|---|---|

| Light content (a handful of points), with instructions | All 7 fit the canvas, with tidy layout and consistent colours |

| Dense content (a dozen figures), thinking on | Plain instructions: all 3 overflowed the canvas. Stressing "must fit": 1 of 2 fit |

| Dense content, thinking off, stressing "must fit" | Both fit; in one the footer slightly overlaps the last card |

| Does it add content of its own? | Yes, e.g. an invented "average value per customer: 1.53", or an unrequested "suggest bringing in contractors". Less often with thinking off, but not gone |

| Content only, no instructions | No slide: it answers like a chat assistant (1 run) |

| Speed | ~30–38 tok/s; 45–90 s per slide with thinking off, 33–270 s with it on |

So in 42model we turn thinking off by default and send a system prompt automatically (fit a 1280×720 canvas, add or drop nothing).

Check figures and facts against your source before showing the slide to anyone. Each row above rests on 2–3 runs: enough to show something happens, not a success rate.

Limitations

• One slide per run. Upstream notes that real use pairs the model with an agent framework (multiple tool calls, re-rendering and revising). We generate once, so results fall short of the numbers in the upstream paper.
• The output is HTML, not a PowerPoint file.
• Needs a machine with about 32 GB of memory (the model file is 21.2 GB).

Use

Packaged for 42model: in the desktop app, open Writing, switch to Slides, and download Logics-PPT when prompted; or run 42model download logics-ppt.

The system prompt, thinking off, and an 8192-token output cap ship with the model, so 42model run logics-ppt < content.txt works as is.

If you call the model from other tools, send this system prompt, otherwise it will not produce a slide:

你是幻灯片设计助手。把用户给出的内容做成一页 16:9 的 HTML 幻灯片（画布固定 1280×720 像素）。要求：所有内容必须完整显示在画布内，不得溢出或被裁切；元素不重叠；版面平衡；字体清晰可读；避免大片空白。内容较多时，优先缩小字号、减少装饰，不要删减内容，也不要添加原文没有的信息。只输出完整的 HTML 代码。

(It is in Chinese because that is the wording we tested; we have not tested an English translation of it.)

Files and license

| File | Role | Size |

|---|---|---:|

| Logics-PPT-Qwen3.6-35B-A3B-RL-Q4_K_M.gguf | the model (Q4_K_M) | 19.7 GiB |

| mmproj-Logics-PPT-Qwen3.6-35B-A3B-RL-f16.gguf | vision encoder (the base is multimodal; not needed for slides) | 0.84 GiB |

Weights are under the upstream Apache-2.0 license.