How to run Emu3.5 locally

BAAI · Any to any world model (gen + edit) · 34B params · Apache-2.0 (commercial OK)

A 34B any-to-any autoregressive 'world model' that predicts next states across vision and language for interleaved image+text generation, image editing and simulation, trained on 10T tokens. The largest and most capable open omni model here, but heavy to run locally.

What it costs to run — $0 markup

On your own machine — $0. Runs free if you have about 48 GB of VRAM (~90GB in BF16; needs 2x 48GB (A6000/L40S) or an 80GB H100 at full precision. 4-bit quantization can bring it toward a single 48GB card.), via Transformers.
Rent a GPU — from $1.39/hr. Fits A100 80GB at the direct vendor price ($0 markup) — pay only for the minutes you generate.
Download the weights — free. Open weights at BAAI/Emu3.5.

Note: generative-media models are billed per image / per second / per minute on hosted services — not per token. Running locally or on your own rented GPU is usually far cheaper and keeps your data on your machine.

Key facts

Does	Text → image, Image editing, Interleaved generation, Visual understanding, World modeling
VRAM to run	~48 GB (~90GB in BF16; needs 2x 48GB (A6000/L40S) or an 80GB H100 at full precision. 4-bit quantization can bring it toward a single 48GB card.)
Download	~90 GB
Parameters	34B
License	Apache-2.0 (commercial use OK)
Run with	Transformers

Get Emu3.5 on Hugging Face →

More multimodal / omni models

Qwen2.5-Omni-7B — Omni understanding + speech generation, ~24 GB VRAM
Janus-Pro-7B — Unified understanding + image gen, ~16 GB VRAM
OmniGen2 — Any to image (gen + edit + in context), ~17 GB VRAM
MiniCPM-o 2.6 — Omni understanding + speech generation, ~9.0 GB VRAM
BAGEL-7B-MoT — Unified understanding + image gen + edit, ~24 GB VRAM
Emu3-Gen — Next token any to any generation, ~18 GB VRAM

Browse: all media models · chat / LLM models