Published: 2026-07-28

A Self-Correcting Diffusion Agent Model From Inclusion AI

Chapters / key moments (click to jump — plays here on the page)

Inclusion AI, part of Ant Group, has released an open-weight diffusion model aimed at agent workloads: instead of predicting one token at a time left to right, it drafts a whole block of text and refines it over successive passes. Two design choices matter for agents — Levenshtein-style editing gives it explicit delete and insert moves so it can rip out a wrong chunk and slot in a correction mid-generation rather than plowing forward with the mistake baked in, and block routing keeps its mixture-of-experts selection efficient over long contexts. The video cites roughly 703 tokens/second on the BFCL tool-use benchmark with a score around 61, close to double the throughput of a comparable model on the same work, with weights on Hugging Face and GitHub. It's around 100B parameters, so it needs real GPU capacity or a cloud host — not a laptop.

Two caveats on this one

First, the video is largely a promotion for the creator's paid community, with the model used as the hook — we've summarised the technical claims and dropped the rest. Second, the auto-generated captions render the model's name inconsistently (as both "Ladder 2.2-Flash" and "Llama 2.2-Flash"), so we've deliberately not stated a product name we can't verify. Check Inclusion AI's Hugging Face and GitHub listings for the actual release name and benchmark figures before relying on any of these numbers.

Watch on YouTube → · Compare agent model benchmarks → · ← Back to News

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Configuration shown

Model facts as stated (unverified; the transcript spells the name "Ladder" and "Llama" 2.2-Flash)

From Inclusion AI (part of Ant Group). A diffusion language model, mixture-of-experts, about 100 billion parameters, 128,000-token context. Open weights on Hugging Face and GitHub. Features named: Levenshtein editing (built-in delete and insert moves so it can remove a wrong chunk and insert a better one while writing) and block routing (keeps expert selection tight over long inputs). Scores quoted: about 49 out of 100 on a coding-bugfix test (name garbled in the captions), about 61 on BFCL tool use, about 703 tokens per second on that tool-use test, close to twice the speed of a comparable model.

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

Who it is for
People building agents that take many tool-using steps, where speed and self-correction matter. For plain chat the speaker says you do not need it.
Run it
Download the weights from Hugging Face or GitHub. He warns it is a roughly 100-billion-parameter model, so plan for real GPU capacity or a cloud service that provides it, not a laptop.
Settings
Start with the defaults recommended on the model page. Pushing the speed-versus-quality control too far toward speed can make output messy and repetitive.

Gotchas

Promotional avatar video for a paid community; no demo run is shown, only descriptions and benchmark figures.
The model name is unclear in the captions, so confirm it against the Hugging Face model page before relying on any figure.
The claim that this is among the first diffusion models pushed to this size as a real agent is the speaker's, not checked.