mere.run

Local text, image, video, music, speech & 3D from one CLI. Native Swift/MLX — Metal on Apple Silicon, CUDA on Linux. No Python at inference. MIT.

[mere.run @meredotrun]

One CLI. Text, image, video, music, speech and 3D — generated on your own machine.

Native Swift on MLX. Metal on Apple Silicon, CUDA on Linux. No Python at inference. MIT licensed.

This demo ran start to finish on one Mac:

[mere.run @meredotrun]

Eight days. Eight releases. 52 pull requests.
https://t.co/hSpxszYFsn v0.43.0 brings local AI to iPhone, adds full LTX 2.5 + MiniMax Music 3, and makes repeat LTX prompts up to 23.8× faster.
More native models. Reproducible evals. More places to run.
https://t.co/YmzGwNnGk8

[mere.run @meredotrun]

@nopmobiel @ivanfioravanti We’ve got an initial MLX implementation running now, optimizing in the works for Mac silicon and CUDA. Give it a try if you get a chance!

https://t.co/HKtSm6Fsl3

[mere.run @meredotrun]

@alexocheema We could use a couple friends over here! https://t.co/Wp8EXm5NiW

[mere.run @meredotrun]

There is a promising future for truth in capabilities and distilling signal from noise when it comes to what works on your box - join up - https://t.co/xErYPrHSJL

[mere.run @meredotrun]

There is a promising future for truth in capabilities and distilling signal from noise when it comes to what works on your box - join up - https://t.co/xErYPrHSJL

[mere.run @meredotrun]

Huge thanks to @antirez. h3.c was an invaluable reference for the serving-parity work behind the 69.9% H3 Ref2VA speedup.
We also ship his DwarfStar (DS4) runtime for local DeepSeek V4 Flash on Metal and CUDA.
Open source compounds. 🙏

[mere.run @meredotrun]

https://t.co/hSpxszYFsn v0.36.0 is out.
Yesterday’s model drops. Local today.
LTX 2.5 · Nemotron + DSpark · Muse Glimmer + DFlash · H3 Ref2VA up to 69.9% faster · ACE-Step improvements
Apple Silicon + CUDA.
MLX just keeps getting better.
https://t.co/GU26dLd7OK

[Rubén @enrub_en]

Mix things, push boundaries

[mere.run @meredotrun]

The whole surface on one sheet, as of v0.35.1: 131 commands, 98 managed models, 21 model categories, 3 platform lanes.

Numbers pulled straight from the CLI — commands walked from --help, model counts from status, 94 of the 131 pinned in the typed contract at catalog --json.

[mere.run @meredotrun]

The launch film. Every asset was generated locally by an agent driving the CLI on one Mac — the narration is a TTS voice clone, the score is ACE-Step, the SFX are Woosh, the stills are Z-Image, and the closing shot is a SCAIL-2 subject swap. The film is the receipt.

[mere.run @meredotrun]

One CLI. Text, image, video, music, speech and 3D — generated on your own machine.

Native Swift on MLX. Metal on Apple Silicon, CUDA on Linux. No Python at inference. MIT licensed.

This demo ran start to finish on one Mac:

[mere.run @meredotrun]

MIT licensed. One-line install: macOS DMG, Linux x86_64 + arm64 CUDA packages (validated on a DGX Spark).

https://t.co/K4UojDXUs2
https://t.co/rrUkjfJuZg

[mere.run @meredotrun]

One CLI. Text, image, video, music, speech and 3D — generated on your own machine.

Native Swift on MLX. Metal on Apple Silicon, CUDA on Linux. No Python at inference. MIT licensed.

This demo ran start to finish on one Mac:

[mere.run @meredotrun]

Local video with its own soundtrack — picture and audio generated together, then muxed to MP4. A few minutes on a laptop, no cloud in the loop.