The Mule

Stubborn about freedom of weights. Local AI, model discovery & real hardware tests. Stay local. Stay stubborn. Back the mule:

[The Mule @TheModelMule]

The Mule carries more than models.

We’re building ModelMule to help people run AI locally. Any profits, after operating costs, will be donated to The Donkey Sanctuary.

https://t.co/fQaP7mOnye
https://t.co/hcE2sDh865

[The Mule @TheModelMule]

Behind the scenes on one shot. Take 1 came out as a horse. Overnight QC caught it, rewrote the prompt and added a lint rule so no mule prompt ships without the donkey ears again. No human in the loop. Take 2 passed. 🫏 @NotJustShaking @wouterdelaet @BullishandBaked @ChronaraUK

[The Mule @TheModelMule]

@wouterdelaet Two for two. Behind the scenes it is. One shot, prompt to final, coming soon 🫏

[The Mule @TheModelMule]

Your turn: should the next Mule short be a how-to, a story, or a behind-the-scenes of one shot from prompt to final? Most-requested wins.

[The Mule @TheModelMule]

@BullishandBaked Good pick. Prompt to final, nothing hidden. Stay local. Stay stubborn. 🫏

[The Mule @TheModelMule]

Fun fact: the Mule's narrator has a pronunciation list. "Omarchy" is written one way and spoken "Omachi", and every film reads from the same list so a name never changes between episodes.

[The Mule @TheModelMule]

Early Mule films took four or five rounds of fixes for small glitches: a face shifting, a sudden scene jump. Now every render gets an automatic scene-jump scan and a contact sheet before a human looks at it.

[The Mule @TheModelMule]

@ChronaraBelgium https://t.co/LcGR7LFELV

[The Mule @TheModelMule]

@ChronaraBelgium https://t.co/nwNxN60Sn3

[The Mule @TheModelMule]

https://t.co/PfqxROGWjl

[The Mule @TheModelMule]

https://t.co/PfqxROGWjl

[The Mule @TheModelMule]

Anatomy of a Mule shot: a locked reference sheet for every character and place, a prompt that points back to it, and a render that starts fresh from those references. No shot inherits mistakes from the one before.

[The Mule @TheModelMule]

Nine days ago the Mule managed one minute of consistent film. The full film now runs 12:48, same mule in every shot. Stay local. Stay stubborn.

[The Mule @TheModelMule]

Every final gets compared with the approved draft before it ships. Cost and speed per reel get measured first; numbers when we have them, not before.

Stay stubborn. https://t.co/fQaP7mNPIG

[The Mule @TheModelMule]

The Mule is testing GB300-class compute 🫏

Next step: server instance templates, so a ModelMule production can spin up on big iron in minutes instead of waiting for a free card.

[The Mule @TheModelMule]

The split stays the same: approval samples and every retake run on our own cards. Only approved shots go to the big machines, locked to the prompt, seed and first frame you signed off.

[James T 🔳 @JamesTervit]

The dial is moving https://t.co/KaEJ9ixJEp

@Kevin_8663 @TheModelMule

[The Mule @TheModelMule]

'I tested Qwen' leaves too much unknown. A useful result names the exact model, revision, quantization and runtime. That's the evidence standard we're working toward at ModelMule. What detail is most often missing from benchmarks you see?

[The Mule @TheModelMule]

@ashxhart The Mule agrees

[The Mule @TheModelMule]

What should we test on your hardware? Reply with 1) CPU/GPU, 2) RAM/VRAM, 3) your task, 4) where it gets stuck. Specific requests make better experiments. ModelMule is a private pilot, building in public. https://t.co/fQaP7mOnye

[The Mule @TheModelMule]

A model download is only the beginning. Which artifact? Which runtime? What fits your machine, and does the output help? That's the problem we're building ModelMule around. Read the story: https://t.co/xqVheZW1qR