LTX 2.5 VRAM Requirements
No official LTX 2.5 checkpoint fits 16 GB. The smallest official transformer plus the smallest official Gemma 4 text encoder already add up to roughly 34 GB. 16 GB cards need the community GGUF transformer and GGUF Gemma 4 encoder instead — this page has the full quant table.
Last updated: 2026-08-26
Quick answer by VRAM tier
| VRAM | Path | Notes |
|---|---|---|
| 16 GB (RTX 4060 Ti, 4070, 4080) | GGUF only | No official combination fits — use the community GGUF transformer + GGUF Gemma 4 encoder below |
| 24 GB (RTX 3090, 4090) | Official INT8/NVFP4, or GGUF Q6_K/Q8_0 | Smallest official combo (NVFP4 + INT8 Gemma 4) is tight at 24GB; GGUF Q6_K/Q8_0 leaves more headroom |
| 32 GB (RTX 5090) | Official INT8/NVFP4 comfortably, BF16 with offloading | NVFP4 gets native matmul speedups on Blackwell |
| 48 GB+ (A6000, A100, H100) | Official BF16 full precision | Full transformer + full Gemma 4 encoder resident, no offloading needed |
Why official files don't fit 16 GB (or even 24 GB, comfortably)
Every LTX 2.5 workflow needs one transformer plus one Gemma 4 text encoder. Here's what the smallest official combinations actually add up to.
| Transformer | Text encoder | Combined | Practical min VRAM |
|---|---|---|---|
| NVFP4 distilled — 18.72 GB | INT8 convrot Gemma 4 — 15.37 GB | ≈ 34.09 GB | 24 GB (tight) / 32 GB comfortable |
| INT8 convrot distilled — 21.50 GB | INT8 convrot Gemma 4 — 15.37 GB | ≈ 36.87 GB | 32 GB |
| BF16 distilled — 42.02 GB | BF16 Gemma 4 — 26.26 GB | ≈ 68.28 GB | 48 GB, or 32 GB with heavy offloading |
Figures are the sum of the two required files' on-disk sizes, not a measured runtime footprint — ComfyUI offloading changes actual peak usage, but the gap to 16GB (or even 24GB) is large enough that offloading alone doesn't close it. This is why the official int8-convrot checkpoints are tagged for 24GB+ on this site rather than 16GB.
The community GGUF path (16 GB+)
Not gated, and small enough for 16 GB when paired with the GGUF Gemma 4 encoder. Requires the ComfyUI-GGUF custom node to load. Rows with a page on this site link to their download page; the rest link straight to HuggingFace.
| File | Size | Repo |
|---|---|---|
LTX-2.5-Distilled-Q3_K_S.gguf | 12.65 GB | Abiray/LTX-2.5-Distilled-GGUF |
LTX-2.5-Distilled-Q3_K_M.gguf | 12.92 GB | Abiray/LTX-2.5-Distilled-GGUF |
LTX-2.5-Distilled-Q4_K_S.gguf | 15.33 GB | Abiray/LTX-2.5-Distilled-GGUF |
LTX-2.5-Distilled-Q4_K_M.gguf | 15.69 GB | Abiray/LTX-2.5-Distilled-GGUF |
LTX-2.5-Distilled-Q5_K_M.gguf | 18.12 GB | Abiray/LTX-2.5-Distilled-GGUF |
LTX-2.5-Distilled-Q6_K.gguf | 18.62 GB | Abiray/LTX-2.5-Distilled-GGUF |
LTX-2.5-Distilled-Q8_0.gguf | 23.60 GB | Abiray/LTX-2.5-Distilled-GGUF |
gemma4-12b-with-proj-ltx-2.5-Q5_K_M.gguf | 9.51 GB | elix3r/gemma4-12b-with-proj-ltx-2.5-GGUF |
Two more community repos publish LTX 2.5 GGUF quants we haven't built individual pages for: realrebelai/LTX-2.5_GGUFs (includes a Q2_K at 8.83 GB, the smallest known LTX 2.5 quant) and vantagewithai/LTX-2.5-GGUF (dev and distilled quant sets). Check each repo's file listing directly for the current set of quants.
Frequently asked questions
What is the minimum VRAM to run LTX 2.5?▼
16 GB VRAM, but only via community GGUF quants — the LTX-2.5-Distilled-Q3_K_S/Q4_K_M GGUF transformer plus the gemma4-12b-with-proj-ltx-2.5-Q5_K_M GGUF text encoder (9.51 GB). None of the official Lightricks/LTX-2.5 files fit a 16 GB card once you add the required Gemma 4 encoder.
Why doesn't LTX 2.5 fit on 16GB like LTX 2.3 does?▼
LTX 2.5's smallest official text encoder (Gemma 4 12B INT8 convrot) is 15.37 GB, versus 9.5 GB for LTX 2.3's Gemma 3 FP4 encoder. Paired with the smallest official transformer (18.72 GB NVFP4 or 21.50 GB INT8 convrot), the combined official minimum is roughly 34 GB — well past 16 GB, and past 24 GB too.
Do I need to sign in to HuggingFace to download LTX 2.5?▼
For the official files, yes. The Lightricks/LTX-2.5 repo is gated: you must be signed in to HuggingFace and click "Agree and Access" before any download link works. The community GGUF quants (Abiray, elix3r) are not gated.
What's the recommended LTX 2.5 setup for 24GB VRAM?▼
The official NVFP4 distilled transformer (18.72 GB) plus the official INT8 convrot Gemma 4 encoder (15.37 GB) is the smallest all-official combination, though it's tight at exactly 24 GB. The community GGUF Q6_K (18.62 GB) or Q8_0 (23.60 GB) transformer paired with the GGUF Gemma 4 encoder leaves more headroom.