Can the RTX 3090 24 GB run LTX-2 (19B)?
Not at full quality. The 16-bit and 8-bit files are too big for the RTX 3090 24 GB; the largest that fits is Q6_K (16.0 GB).
Take Q6_K (16.0 GB). In use it comes to about 20.8 GB, with 3.2 GB left over on 24 GB. Quality: very close to the original. For better quality, Q8_0 (20.4 GB) also runs, with about 1.2 GB spilling into system RAM - a little slower, still practical.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | ltx-2-19b-dev-Q6_K.gguf Q6_K · unsloth/LTX-2-GGUF | models/unet | 16.0 GB | Download → |
| Text encoder | gemma_3_12B_it_fp8_scaled.safetensors Gemma 3 12B FP8 (scaled) · Comfy-Org/ltx-2 | models/text_encoders | 13.2 GB | Download → |
| VAE | ltx-2-19b-dev_video_vae.safetensors LTX-2 video VAE · unsloth/LTX-2-GGUF | models/vae | 2.4 GB | Download → |
| Also needed | ltx-2-19b-dev_audio_vae.safetensors LTX-2 audio VAE · unsloth/LTX-2-GGUF | models/vae | 0.2 GB | Download → |
| Also needed | ltx-2-19b-dev_embeddings_connectors.safetensors LTX-2 text embeddings connectors (projection) · unsloth/LTX-2-GGUF | models/text_encoders | 2.9 GB | Download → |
| Total download · keep about the same free on disk | 34.7 GB | |||
Alternatives: Gemma 3 12B FP4 mixed (9.4 GB); Gemma 3 12B BF16 (24.4 GB). FP4 files are listed only as alternatives unless the GPU is an RTX 50 (Blackwell). Only for some workflows: LTX-2 spatial upscaler x2 (1.0 GB, official two-stage template). Sizes from Hugging Face, 2026-10-01.
System RAM: 48 GB or more. These files take about 34.7 GB of RAM while loaded, 40.7 GB with Windows and ComfyUI. What happens with less. calculated
02Every LTX-2 file on 24 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit GGUF · unsloth | 37.8 GB | 42.6 GB | 18.6 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · unsloth | 20.4 GB | 25.2 GB | spills 1.2 GB | practically identical to the original | Hugging Face → |
| Q6_K ← GGUF · unsloth | 16.0 GB | 20.8 GB | fits · 3.2 GB spare | very close to the original | Hugging Face → |
| Q5_K_M GGUF · unsloth | 14.3 GB | 19.1 GB | fits · 4.9 GB spare | close; small differences in fine detail | Hugging Face → |
| Q4_K_M GGUF · unsloth | 12.8 GB | 17.6 GB | fits · 6.4 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · unsloth | 10.1 GB | 14.9 GB | fits · 9.1 GB spare | noticeable loss of detail | Hugging Face → |
| Q2_K GGUF · unsloth | 8.1 GB | 12.9 GB | fits · 11.1 GB spare | heavy loss; a last resort | Hugging Face → |
“Needed” adds 4 GB of working memory for LTX-2 and a 0.8 GB reserve. Sizes 2026-10-01.
03Measured and reported results
Nobody has sent measured numbers for this pair yet. If you run LTX-2 on a RTX 3090 24 GB, send your time per image and peak VRAM and it will appear here, credited.
04With more VRAM
With 32 GB you could run LTX-2 at 8-bit or better (Q8_0, 20.4 GB) with no offloading: for example on the RTX 5090 32 GB.
What changes from the RTX 3090 24 GB to the RTX 4090 24 GB → · What changes from the RTX 3090 24 GB to the RTX 5090 32 GB →
05About the RTX 3090 24 GB
The RTX 3090 24 GB (Ampere) has no FP8 compute. FP8 files still load and save memory, but ComfyUI converts them back before the maths, so there is no speed gain - Q8_0 GGUF is the closer-to-original 8-bit choice. ComfyUI can compute INT8 files natively on NVIDIA, which is worth a try where a model offers one.
24 GB GDDR6X · 384-bit · 936 GB/s · Ampere. Everything that runs on it →
06Questions
How much VRAM does LTX-2 need?
Around 12.9 GB with the smallest file (Q2_K) and 25.2 GB with an 8-bit file (Q8_0). The full 16-bit file needs about 42.6 GB.
Which LTX-2 file should I download for the RTX 3090 24 GB?
Q6_K (16.0 GB) from unsloth/LTX-2-GGUF. Load it with Unet Loader (GGUF) from the ComfyUI-GGUF node pack.
How much do I need to download for LTX-2 on the RTX 3090 24 GB?
34.7 GB: ltx-2-19b-dev-Q6_K.gguf, gemma_3_12B_it_fp8_scaled.safetensors, ltx-2-19b-dev_video_vae.safetensors, ltx-2-19b-dev_audio_vae.safetensors, ltx-2-19b-dev_embeddings_connectors.safetensors.