Best Flux Model for 24GB VRAM (24GB VRAM)
Running the full FP16 Flux.1 reference model on 24GB
With 24GB VRAM you can finally run Flux.1 dev at full FP16 — the reference model, ~24GB, ~12 seconds per 1024px image on an RTX 4090, with no quality compromise. 24GB also unlocks Flux LoRA training and heavy ControlNet + multi-LoRA stacks. If you want a little headroom, FP8 frees ~12GB for higher resolution and bigger batches. This is the tier for serious, no-compromise local Flux work.
Models that fit in 24GB VRAM
Tested reference: RTX 4090 / RTX 3090 (24GB). Figures are for a 1024px image.
| Model | Size | Build | VRAM used | Speed |
|---|---|---|---|---|
| WINNERFlux.1 dev FP16 The full reference model — maximum quality, no quantization. Keep other GPU apps closed. ComfyUI flux1-dev.safetensors (FP16) | 12B | FP16 | ~24GB | ~12s / image |
| Flux.1 dev FP8 (+ heavy LoRAs) Frees VRAM for multi-LoRA + ControlNet + 2048px renders and big batches. ComfyUI FP8 checkpoint | 12B | FP8 | ~12–16GB | ~10s / image |
| Flux.1 dev (LoRA training) 24GB is the practical minimum for comfortable Flux LoRA training (ai-toolkit / kohya). ai-toolkit / kohya_ss | 12B | FP8 base | ~20–24GB | training |
What won't fit in 24GB
- ✗Flux.1 dev FP16 + full fine-tune (needs 40GB+) — Full-model fine-tuning (not LoRA) needs an A100/H100-class card; LoRA training is fine on 24GB.
How to fit more in 24GB
- →24GB runs full FP16 Flux — the only reason to drop to FP8 here is to add LoRAs/ControlNet or push past 1536px.
- →It is the entry tier for training your own Flux LoRAs locally.
- →On a 3090, FP16 Flux works but is ~1.7× slower than a 4090 — still very usable.
Quick start
git clone https://github.com/comfyanonymous/ComfyUIDrop flux1-dev.safetensors into models/unetGo from “it runs” to actually building
Learn to run local models, wire up RAG, and build agents — hands-on, with a free first chapter of every course.
Start learning free →Frequently asked questions
Can 24GB VRAM run full Flux?
Yes — 24GB (RTX 4090/3090) runs Flux.1 dev at full FP16 (~24GB, ~12s/image) with no quality compromise. It is the first tier where you get the reference model rather than a quantized build.
Can I train Flux LoRAs on 24GB?
Yes — 24GB is the practical minimum for comfortable local Flux LoRA training with ai-toolkit or kohya (FP8 base + gradient checkpointing), at 512–1024px.
Do I need FP16 or is FP8 fine on 24GB?
FP8 is visually near-identical for almost all prompts and frees VRAM for LoRAs/ControlNet/higher resolution. Use FP16 when you want the absolute reference output or are comparing quality.
Related guides
Ready to Go Beyond Tutorials?
20 structured courses with hands-on chapters - build RAG chatbots, AI agents, and ML pipelines on your own hardware.
Was this helpful?