What it is
LTX-2.5 was released on 11 August 2026 as open weights on Hugging Face and through Lightricks' LTX API. It is a roughly 22-billion-parameter diffusion transformer that generates video and audio together from text, image or audio inputs, using a custom Gemma 4 12B text encoder. New in this version are native multishot generation that holds characters and voices across cuts, a diffusion video decoder, a prompt enhancer and a duration head that picks clip length from the prompt.
The release includes a trainable dev checkpoint, a distilled checkpoint on a fixed 8-step schedule, x2 spatial and temporal upscalers, and int8 and NVFP4 builds for ComfyUI. Most LoRAs trained for LTX-2.3 load without changes. Lightricks lists a 32GB NVIDIA GPU as the minimum and an 80GB A100 or H100 as recommended.
On the API, ltx-2-5-fast generates up to 20 seconds at 720p or 1080p and up to 10 seconds at 1440p or 4K; ltx-2-5-pro produces 6, 8 or 10 second clips at every resolution up to 4K. The weights use the LTX-2.x Community Licence: free commercial use for companies under $10 million in annual revenue, counted across affiliates, and a paid agreement above that.
Our review
LTX-2.5 is built for people who intend to run a video model rather than rent one. Day-one ComfyUI templates, a Diffusers port, a distilled 8-step checkpoint, quantised builds and LoRAs that carry over from 2.3 make it far easier to deploy and adapt than its open rivals, and the docs now cover VFX work such as HDR EXR output, alpha mattes and relighting through task-specific IC-LoRA adapters. Lightricks claims a 10-second 720p clip in 6.8 seconds on two Nvidia GB200 GPUs, as reported by LLM Stats; the model card itself publishes no speed figures.
Quality is the trade-off. On Artificial Analysis's text-to-video board with audio, LTX-2.5 Fast ranks 24th at 949 Elo and Pro 25th at 945, level with Veo 3.1 Lite and more than 200 points behind MiniMax H3 at 1158, which is also open-weight. Pro scoring below Fast is the oddity: at $0.17 a second at 1080p against $0.13, the dearer tier buys no measured improvement on that board. Both are a clear step up from LTX-2.3 Pro at 888.
The API has become more expensive. LTX-2.3 Pro costs $0.08 a second at 1080p; 2.5 Fast is $0.13 and 2.5 Pro $0.17, so a minute of 1080p Pro is $10.20 against $4.80 for MiniMax H3. Concurrency defaults to two simultaneous jobs, finished-job records can vanish after 24 hours and output URLs expire, so a production pipeline must copy results to its own storage at once.
The licence is open in practice only for smaller firms. The $10 million threshold counts the whole corporate group, passing on fine-tunes may require a paid licence, and the gated Hugging Face download asks users to consent to targeted advertising. For a studio inside a larger media group, the free weights are a trial rather than a production right.
Where it earns its place
A VFX lead at a 15-person post-production studio
Generate plates, alpha mattes and HDR shots inside an existing ComfyUI pipeline
Run the distilled checkpoint on an 80GB A100 or H100 with the IC-LoRA adapters for mattes, relighting and SDR-to-HDR. There are no per-second fees, and the licence is free while group revenue stays under $10 million.
A product engineer at a social video app
Offer prompt-to-clip generation with sound to users
Call ltx-2-5-fast through the async v2 API; 10,000 ten-second 1080p clips a month is $13,000 at $0.13 a second. The default limit of two concurrent jobs has to be raised through sales before launch.
A marketing manager at a regional retailer
Make short vertical product ads with voice-over
Pro at 1080p in 9:16 costs $1.70 for a 10-second clip, but Fast scores slightly higher on Artificial Analysis for $1.30. Above $10 million in group revenue, the API rather than the free weights is the licensed route.
Before you commit
- Setup
- ComfyUI templates load the model in a few steps on a 32GB+ NVIDIA GPU with CUDA 12.7 and Python 3.12; the API needs a key and an async submit, poll and download loop.
- Integrations
- ComfyUI nodes, Diffusers, native ltx-pipelines in PyTorch, the LTX Trainer for LoRA and full fine-tunes, and the LTX REST API for text, image and audio to video.
- Data and privacy
- Self-hosted weights keep footage in-house. API outputs sit on expiring URLs and job records last up to 24 hours; whether API inputs are used for training is not stated in the docs we read.
- Cost at scale
- 1,000 ten-second 1080p clips a month cost $1,300 on Fast and $1,700 on Pro; 4K Pro is $0.39 a second, or $23.40 a minute. Self-hosting needs 32GB of VRAM at minimum, with an 80GB A100 or H100 recommended.
- Limits
- Fast: up to 20 seconds at 720p or 1080p at 24 or 25 fps, 10 seconds otherwise; Pro: 6, 8 or 10 seconds; two concurrent API jobs by default; no retake or extend on 2.5.
What's good
- Open dev and distilled weights, with an official trainer
- Joint video and audio, including multishot scenes
- API reaches 4K on both tiers; Fast runs to 20 seconds
- Most LTX-2.3 LoRAs carry over
What's not
- 24th on Artificial Analysis's video board, 200+ Elo behind MiniMax H3
- Pro scores below Fast despite costing more
- Free licence ends at $10M group revenue
- 1080p API prices more than double LTX-2.3's
- Self-hosting needs at least a 32GB GPU
Best for
Developers and smaller studios who want an open video model they can fine-tune and run on their own GPUs.
Alternatives
LTX-2.5: common questions
- Is LTX-2.5 worth it?
- Easy to self-host, fine-tune and wire into ComfyUI, LTX-2.5 trades 200 Elo of picture quality for control. We rate it 3.8 out of 5.
- How much does LTX-2.5 cost?
- Weights free under $10M annual revenue (LTX-2.x Community Licence). API per second: Fast $0.09 (720p), $0.13 (1080p), $0.19 (1440p), $0.30 (4K); Pro $0.12, $0.17, $0.25, $0.39.
- What are the downsides of LTX-2.5?
- 24th on Artificial Analysis's video board, 200+ Elo behind MiniMax H3; pro scores below Fast despite costing more; free licence ends at $10M group revenue; 1080p API prices more than double LTX-2.3's; self-hosting needs at least a 32GB GPU.
- Who is LTX-2.5 best for?
- Developers and smaller studios who want an open video model they can fine-tune and run on their own GPUs.
- How much does LTX-2.5 cost for a team or in production?
- 1,000 ten-second 1080p clips a month cost $1,300 on Fast and $1,700 on Pro; 4K Pro is $0.39 a second, or $23.40 a minute. Self-hosting needs 32GB of VRAM at minimum, with an 80GB A100 or H100 recommended.
- Is LTX-2.5 safe for company data?
- Self-hosted weights keep footage in-house. API outputs sit on expiring URLs and job records last up to 24 hours; whether API inputs are used for training is not stated in the docs we read.
- What are the best LTX-2.5 alternatives?
- The closest alternatives we have reviewed are MiniMax H3 (Hailuo 3.0), Wan 3.0, Veo 3.1 and Kling 4.0.
Answers checked Oct 23, 2026.