2 open-weight AI video models: LTX 2.5 and MiniMax H3
Quick answer
The two current open-weight AI video releases to evaluate first are LTX 2.5 and MiniMax H3. LTX 2.5 is the practical starting point when you want a mature public repository, several generation and editing pipelines, and synchronized audio. MiniMax H3 is the stronger second evaluation when multimodal references, native stereo audio, and up to 15-second output matter. This is a deployment shortlist, not a universal quality ranking.
A red ceramic mug sits on a wooden café table beside a window. Steam rises slowly while the camera makes a gentle push-in. Soft morning light, realistic materials, one continuous shot, no people or readable text.
Test a hosted AI video workflow
Run a representative prompt or source image with a current model before deciding whether local deployment is worth the operational cost.
Try AI Video GeneratorIf the open-weight model must run behind a hosted or self-managed endpoint, use the video API and deployment comparison.
LTX 2.5 vs MiniMax H3 at a glance
This comparison uses the official Lightricks repository and MiniMax release documentation checked September 12, 2026. Both publish weights, but “open source” in search language does not mean every component uses an OSI-approved software license. Review the exact model license before commercial deployment.
For LTX 2.5, the LTX-2.x Community License requires a paid license for entities with annual revenue of at least $10 million, except for the license’s defined non-commercial purposes. The threshold aggregates entities under common control. A free weight download is not unrestricted commercial permission; check the applicable model terms and dependencies before deployment.
Model | Best reason to evaluate it | Official inputs and outputs | Access and license | Published local-deployment signal |
|---|---|---|---|---|
A current open-weight audio-video family with multiple production pipelines | Text, image, keyframes, video, or audio conditioning; synchronized video and audio | Downloadable gated weights under the LTX-2.x Community License | Official component download is roughly 66 GiB; quantization and CPU/disk offload are documented | |
A current open-weight omni-modal video model with native stereo audio | Text, first/last frames, images, video, and audio references; 4–15 seconds; 24 fps | Complete H3 weights published in task-specific checkpoints; verify the current model license | MiniMax's SGLang example serves a checkpoint across four GPUs; local capacity planning is required |
1. LTX 2.5: the current Lightricks open-weight release
LTX 2.5 is the recommended model in Lightricks' current LTX-2 repository; LTX 2.3 is now listed as legacy. The repository documents text/image-to-video, keyframe interpolation, audio-to-video, video transformation, Retake, HDR, and synchronized audio-video pipelines.
The official quick start downloads a distilled transformer, a fine-tuned Gemma 4 text encoder, video and audio VAEs, and a spatial upscaler totaling roughly 66 GiB. Production-quality DFR adds a detailing IC-LoRA. Treat checkpoint storage, VRAM, decoding backend, offload settings, and generation time as part of the deployment cost.
Use the official LTX-2 repository and record the exact 2.5 checkpoint, pipeline, commit, license, input files, prompt, dimensions, frame count, and hardware for every retained result.
2. MiniMax H3: open weights with multimodal references and stereo audio
MiniMax open-sourced H3 on August 3, 2026 and says it released complete model weights for further development and fine-tuning. The official release describes two task-specific checkpoints: FL2VA for text or first/last-frame input, and Ref2VA for combinations of image, video, and audio references.
The documented outputs run 4–15 seconds at 24 fps with 32 kHz stereo audio; the shorter image side defaults to 768 pixels, with a separate regenerate stage for 2K. MiniMax's SGLang serving example uses four GPUs. Do not convert that example into a universal minimum: framework, checkpoint, precision, resolution, duration, and offloading determine the real requirement.
Start with MiniMax's official H3 release and deployment guide. Confirm the current weight license and every dependency before commercial use.
How to compare open video models without inventing a winner
Run the same rights-cleared brief through the exact checkpoints and retain every attempt. Separate visual quality from operational fit: a model that can make one strong clip may still fail the throughput, reproducibility, license, or hardware requirements of a production system.
Test | Fixed input | Pass condition |
|---|---|---|
Text to video | One prompt with a subject, action, setting, camera move, and duration | The requested subject and action persist without a disqualifying artifact |
Image to video | One rights-cleared source image and restrained motion instruction | Identity, shape, labels, and important text remain usable |
Video to video | One short rights-cleared source clip and one bounded edit | Timing and unedited regions remain stable while the requested change is visible |
Operations | The exact deployment configuration | Runtime, peak memory, failure rate, and accepted-output cost meet the project limit |
Rights | Code, weights, inputs, outputs, and hosting path | Every license and permission is recorded for the intended territory and use |
What to record for a reproducible evaluation
- Repository commit, checkpoint filenames, license files, inference framework, and dependency versions.
- Prompt and every image, video, or audio reference, plus seed and all generation parameters.
- GPU model and count, VRAM, CPU RAM, storage, precision, offloading, runtime, and peak memory.
- Every attempt, failure, accepted output, and the acceptance rule applied before viewing results.
- Compute cost per accepted clip and the rights for code, weights, inputs, outputs, and distribution territory.
Should you self-host or use an API?
Self-host when model-level control, private infrastructure, or predictable high utilization justifies the engineering and hardware. Use a managed API when fast integration, burst capacity, and avoiding model operations matter more. For an API catalog with many current video endpoints, evaluate fal.ai's video APIs; for a unified creation API and editing tools, compare the Magic Hour AI video API.
Frequently asked questions
Evaluate LTX 2.5 first for its current public repository and broad audio-video pipeline coverage, then MiniMax H3 when multimodal references and native stereo audio are central. Choose only after both pass the same license, hardware, output, and accepted-cost tests.
Lightricks now recommends LTX 2.5 and labels LTX 2.3 as legacy in the current repository. Existing 2.3 integrations can still run, but new evaluations should begin with 2.5 unless compatibility requires the older checkpoint.
MiniMax calls H3 open source and publishes complete model weights plus deployment instructions. For procurement, inspect the current weight license and dependencies directly; a public repository or downloadable checkpoint alone does not settle every commercial-use question.
Downloading weights can avoid a per-generation API fee, but storage, GPUs, electricity or cloud compute, engineering time, retries, monitoring, and output review remain real costs. Compare cost per accepted clip, not download price.
