i have only 16gb vram so i guess at q5 is max i can use ?

@pink0909 With recent comfyUI managing memory far more efficiently, you might even be able to run the Q8 with zero problems (it also can unload on normal ram) so if you have 32 or 64 or more GB ram in addition you can try, but sure the Q5 should be zero problem, I have 24GB Vram and I use the Q5 UD usually, and I never seen an OOM so far with any LTX 2.3 even at 500+ frames, (I recommend the Q5 by unsloth, with the UD version, slightly higher size than usual Q5 but better quality)

I have 12gb vram and tried Q2, got very 'muddy' videos and can't solve it. Is is what you mean by 'dont even bother with q2...'?

@fakolonya yeah, q2 is basically 2 bit version of a 16 or 32 bit original model, too much loss, You probably can run Q4, and preferably the Q4 version made by unsloth (the UD version) in comfyUI unless you only have 16GB of ram in this case it might be a bit tight, because comfyUI can unload dynamically to ram to not go oom usually

@fakolonya Yeah Q2 is just too bad, Q3 UD might be decent enough, and Q4/Q5 sweet spot, yu should be able to run it, maybe reduce some resolution and frame lenght if you encounter oom.

@darksidewalker How I'm waiting for your model, my friend)

Which model did you use? Q4, Q5, Q3?

just throse these errors at grok ai and claude ai. usually they point in the right direction.Register with free mail so you dont get limited to 5 questions. They are better than chatgpt for m my months of experinece.

@RalFinger ah ok , i know qwen only from local uncensored(abliterated) LLMs i load from huggingface. i will checkt this one out how its online.

It isn't a curse, it's taking the time to learn

@pink0909 "So you don't get limited to 5 questions" Can you elaborate? I sign in with my google account and I'm limited to 5 questions, or something like that.

@JLH3 some months ago rule of thumb was the usual companies (chatgpt,grok,claude let you ask 3-5 question as guest , then demand that you login.From there you could ask many questions for hours without getting limited. just recently since about a week, they all seem to get very restricive (you might noticed when grok service broke down with warning they had "issue" (too much traffic i guess). Since now Grok claims to be busy after 1-2 Questions. Claude gives me after 5-7 Question a time limit and says i can come back in 5 hours. I just hop to other Websites like Grok,Chatgpt , and Qwen.ai. usually that does the trick. Also i am testing a lot of local models and found an interesting workaround. Initially i wanted to help a friend who doesnt have a gpu, so i tried to find a way how you can use LLMs with cpu. First i figured out that CPU's can handle max about 4B(illion) parameter models when they get too slow (about 9token/s) , but 4B is to small to be "intelligent" (which starts about 7-9B from my experience (there are exeptions offcourse). But then i found that there exist MOE (mixture of experts) Models which combine a lot of small models to be able to be more capable.The Names usually are like 24B-A4 meaning 24B Paramter (which is a lot) but each expert is just 4B. So here comes the trick: this big 24B model will run on your cpu with the same speed as a small 4B model. There are even 3B Models (about 14token/s) and very fast 2B Model. Just to get you started,all you need is 24gb ram or better, you basicly need only 2 things a gui to load a gguf file my favorite is https://github.com/oobabooga/textgen/releases , for the files i drop some names, https://huggingface.co/mradermacher/Nemotron-Cascade-2-30B-A3B-heretic-GGUF/blob/main/Nemotron-Cascade-2-30B-A3B-heretic.Q4_K_M.gguf if you just want to test it here is a smal MOE, 4b in total but only 2B active meaning it will be fast on weak cpus great for testing https://huggingface.co/mradermacher/Gemma4_E2B_Abliterated_Baked_HF_Ready-GGUF/blob/main/Gemma4_E2B_Abliterated_Baked_HF_Ready.Q4_K_M.gguf let me know how it goes

@pink0909 Thanks for the informative reply. I forgot to mention that I was specifically talking about Claude in my previous comment. There are free models like Kimi and Gemini that will talk to you all day in "fast" mode.

@JLH3 yes claude and grok limit even when logged in, but i noticed yesterday that in night time claude did not limit me i had a long conversation like 15 questions or so without final limit. Also if you want an even easier etry into local LLM there is https://lmstudio.ai/ with built in downloader, the downside is your will not find all models from hugging face, but as a start its good an later you could switch to webui text obongo i mentioned earlier. Just remember for CPU's a Model should not be larger than 4B paramters and smaller means faster (like 3B,2b,1,7b) and best intelligence are MOE-variants, Models with 1,7b-4B Experts in it.

Assuming the VAE OOM's before the saved video, it's due to RAM being full. The only solutions are upgrading to 64GB of RAM or you can lower your resolution to 512x512 and 3 seconds. You can experiment with those 2 settings. Even with a 5090 and 64GB of RAM, I still run out of RAM with the fp8 base model, it happens but less often. You can also lower your monitors resolution in windows to free about 500MB of VRAM (Be careful as your desktop icons will get rearranged), windows uses about 1.2GB VRAM on 4K monitors, you can save another 2GB of system RAM by switching to Linux OS. Steam also uses 500MB of RAM in the background. RBG software can also use 250MB of RAM too depending on what you use. At 32GB you basically have to run the AI program and that's it. Anything that opens or runs in the background can make the difference between it works or not working, that includes windows deciding it wants to update. The only real solution is to upgrade to 64GB of RAM. LTX2 for me, is more memory friendly than wan2.2, at the end of the day, any video model, I still OOM.

Try cachedit ltx-2 and ltx2 attention tuner