Skip to content

Otto365 AI Brief #041

Hugging Face: IBM's small model builds decks in your design language

How this was made: Otto365 and Sam are animated characters and their voices are generated by AI. Sam is a fictional small-business owner. The video tells the same story as the text on this page; the transcript lists every spoken word and the main lines shown on screen. Spotted an error?

Published 4 Nov 2026, 11:30 UK
By the Otto365 AI Brief. The script was drafted with AI assistance, and every claim was checked against the sources below by our automated checks. Published on 4 Nov 2026 under the standing approval of Peter Ferguson, Editor.
Announced 3 Nov 2026
Voiced by AI

What happened

On Hugging Face, IBM's researchers show a small model building on-brand decks. IBM Research announced Palette on the 3rd of November 2026 on Hugging Face: a 20B open-weight model fine-tuned to build decks. A designer plans the deck and a coder draws each slide.1

Three things to know

Designer and coder First, the designer writes a JSON brief: palette, type system, and each slide's layout and content. The coder writes each slide's code in parallel, about 6 to 7 times faster, IBM Research says.1
Directives, then a check Second, plain-English directives like make this a timeline or this is for the board steer the design. A harness renders every slide, flags overlaps and overflow, and keeps a rewrite only if the checker confirms it improved.1
Benchmark: second Third, on PresentBench it placed second overall, 1.1 points behind NotebookLM, IBM Research says. Frontier models drift back to their own house style, and confidential content sometimes can't leave your systems.1

Why it matters to you

Otto365's take (opinion)

So, my take. IBM's lesson: fixing example decks changed the output more than adding rules. For a UK small business, give your AI tool one on-brand slide to copy, and check every slide for invented facts.1

Do this week

Try this: hand your AI deck tool one real slide as its example.

Sources

IBM Research: Palette: teaching a small model to build decks in your design language (opens in a new tab) Primary source, accessed 4 Nov 2026

The quotes we checked (18)

Both of those point to a model you own: open weights, running on your own infrastructure, trained on your own design language.
IBM Research · source 1
That's fine until you want the deck to look like your template.
IBM Research · source 1
We fine-tuned a 20B model to build decks the way people do: a designer plans the deck and a coder draws each slide.
IBM Research · source 1
Enterprise Article Published November 3, 2026
IBM Research · source 1
The designer reads the plan and writes a structured JSON brief: the palette, the type system, and for each slide its layout, its exact content and a short design rationale.
IBM Research · source 1
the coder calls run in parallel, one per slide, which makes the coding stage about 6-7× faster.
IBM Research · source 1
The coder takes one slide's spec from the brief and writes the pptxgenjs JavaScript that draws it, preceded by its reasoning.
IBM Research · source 1
Palette lets them ask for these with directives, short instructions written directly in the plan and wrapped in [[ ]]:
IBM Research · source 1
Without the directive, the model would reasonably show these milestones as a table; with it, the model draws a timeline instead.
IBM Research · source 1
Audience [[this is for the board]]
IBM Research · source 1
Then each slide is rendered and measured: a geometry checker flags text that overlaps, gets clipped or runs off the slide.
IBM Research · source 1
Flagged slides go to a separate editor model, which rewrites that slide's code.
IBM Research · source 1
The key rule is that a rewrite is kept only if the checker confirms the slide improved; otherwise the original stays.
IBM Research · source 1
Across the three domains it places second, 1.1 points behind NotebookLM (59.7 vs 60.8) and ahead of the rest of the field, as a single 20B open-weight model.
IBM Research · source 1
Frontier models have their own house style, and no amount of prompting stops them drifting back to it. And sometimes the content is confidential and can't leave your systems at all.
IBM Research · source 1
The teacher copied its reference decks far more than it followed instructions, so fixing the reference was what actually changed its output.
IBM Research · source 1
When the teacher got something wrong, our first instinct was to add a rule to its prompt, and that rarely helped.
IBM Research · source 1
The worst failure is a slide with facts that weren't in the input, and we found that a prompt rule can't prevent it.
IBM Research · source 1

Full transcript

Sam: Otto! A small model that builds decks in your own style!

Play from 0:04 Otto365: I know. IBM Research announced Palette on the 3rd of November. Sam: What does it actually do? Otto365: A 20B open-weight model, fine-tuned to build decks. A designer plans it and a coder draws each slide.

Play from 0:17 Sam: So who decides the colours? Otto365: The designer writes a JSON brief: palette, type system, and each slide's layout and content. Sam: And is it quick? Otto365: The coder writes each slide's code in parallel. IBM Research says that's about 6 to 7 times faster.

Play from 0:33 Sam: Can I tell it what I want? Otto365: In plain English. Directives like make this a timeline, or this is for the board, steer the design. Sam: What stops text running off the slide? Otto365: A harness flags overlaps and overflow. It keeps a rewrite only if the checker confirms it improved.

Play from 0:50 Sam: Is it actually any good? Otto365: On PresentBench it placed second, one point one behind NotebookLM, IBM Research says. Sam: Why not just use a big model? Otto365: Frontier models drift back to their own house style. And confidential content can't always leave your systems.

Play from 1:06 Sam: So I write a big list of brand rules! Otto365: IBM found fixing example decks changed the output more than adding rules. My take: give your AI one on-brand slide to copy, and check every slide for invented facts. Sam: Goodbye, brand guide.

Play from 1:23 Otto365: It had a good run. Try this: hand your AI one real slide as its example.

Otto365: I'm Otto365.

Follow the Brief.

On screen

Play from 0:00: Otto365 AI Brief #041, Hugging Face, 3 Nov 2026; A small model that builds decks in your own style; Your plan; Designer brief; Coder per slide; Harness check
Play from 0:04: Palette: on-brand decks; Announced 3 Nov 2026; Palette, a 20B open-weight model; Designer plans the deck; Coder draws each slide; Harness checks every slide; Source: IBM Research, 3 Nov 2026
Play from 0:17: Designer and coder; Your plan; JSON brief; Code per slide; Rendered slides; Brief: palette, type, layout; Coder runs 6 to 7 times faster
Play from 0:33: Directives, then a check; Without: Shown as a table; With directive: Make it a timeline, Drawn as timeline; Render slide; Flag overlaps; Editor rewrites; Keep if improved
Play from 0:50: Benchmark: second; NotebookLM: Mean 60.8; Palette: Mean 59.7, 20B open-weight; Frontier model: Its house style, Content sent out; Model you own: Your own style, Own infrastructure
Play from 1:06: Show it one real slide; Add a rule: Rarely helped; Fix the example: Changed the output; One real slide as the example; Check facts against your input
Play from 1:23: Give it one real slide; Voiced by AI. Fact-checked against its sources. Apps 365 sells AI and Microsoft 365 services. Not endorsed by companies named.

Voiced by AI. Fact-checked against its sources. Apps 365 sells AI and Microsoft 365 services. Not endorsed by companies named. Report an error.

Add Otto365 as a preferred source on Google