Start with the license and the machine, then choose the model. Wan2.2-TI2V-5B is a useful first candidate for a permissively licensed text-to-video and image-to-video experiment. LTX-2.5, HunyuanVideo-1.5 and MiniMax H3 deserve evaluation for different workflows, but their custom terms and runtime requirements matter before a demo does.
Checked October 31, 2026. The shortlist is source-based; we also generated one small Wan community-conversion preview on a Mac mini, shown below. It is a functional test, not a cross-model video benchmark. Our broader video-model article addresses hosted generation; this guide asks what you can download and operate yourself.
Four candidates, four deployment decisions
| Model repository | What to evaluate | Declared license | Deployment detail to check first |
|---|---|---|---|
| Wan2.2-TI2V-5B | One 5B checkpoint for text-to-video and image-to-video | Apache 2.0 | Official single-GPU example specifies at least 24 GB VRAM with offloading |
| LTX-2.5 | Audio/video generation and conditioning workflows | LTX 2.x Community License | Gated download; verify the exact runtime, checkpoint and full license |
| HunyuanVideo-1.5 | Text/image-to-video; distinct resolution and distilled variants | Tencent Hunyuan Community | Official requirements specify NVIDIA CUDA and a 14 GB minimum with offloading; territorial restrictions apply |
| MiniMax H3 | Reference-conditioned video with generated audio | MiniMax H3 Community | Multiple pipeline components; territorial restrictions apply |
Sources: linked official model cards, metadata and license files as of 2026-10-31. Hardware figures describe the publishers' stated configurations, not measurements by Teamday or interchangeable minimums for every workflow. This is a curated shortlist, not a ranking of every Hub repository.
The licensing difference that changes the shortlist
Open weights describe availability, not one standard set of permissions. A model listed beside an Apache-licensed checkpoint may have a very different license.
The checked HunyuanVideo-1.5 license excludes the European Union, United Kingdom and South Korea from its defined territory. The checked MiniMax H3 license excludes those territories and the United States. These restrictions are a reason to eliminate a candidate from a default deployment plan in those locations unless you obtain appropriate separate rights. Hosted-service access is a separate arrangement; do not assume an API offer grants the same rights as a weight download.
LTX-2.5 also uses a custom community license and has gated access. We verified its license identifier and access status, but did not obtain the gated license file in this evaluation; we make no blanket commercial-permission claim for it. Follow the current agreement presented by its repository.
This distinction is why an article called “free video models” can mislead. Compute costs money, licenses differ, and a downloadable artifact does not necessarily fit your intended business use.
Why parameter count is a poor hardware estimate
A video pipeline includes more than its main denoising weights. Text encoders, latent video, attention buffers and decoding all consume memory. Increasing the number of frames or output resolution can change the peak substantially. CPU offloading reduces GPU residency but moves work and data across the host system.
Wan's official TI2V-5B command uses model offloading, dtype conversion and a text encoder on CPU. The documented at-least-24-GB GPU requirement applies to that reference path. A community quantized workflow may use less GPU memory, but it is a different configuration and may have different speed or output behavior.
HunyuanVideo's card distinguishes 480p, 720p and distilled variants. A speed claim for a distilled image-to-video configuration should not be attached to the whole model family. Record the full variant, step count, resolution and frame count beside every timing you publish.
For a Mac mini, verify an actual Apple-compatible implementation. A code snippet that substitutes mps for cuda does not establish that every kernel, dtype and component works. Unified memory capacity is also not equivalent to the same number of gigabytes of discrete GPU VRAM. Our Mac preview below establishes one community MLX configuration; it does not validate the official CUDA path on MPS or establish a minimum memory requirement.
Our Mac mini preview: Wan through a community MLX conversion
We made one original text-to-video attempt on October 31, 2026: Apple M6 Mac mini, 32 GB unified memory, macOS 27.0.1. The runtime was MLX-Gen 0.38.0 (its mflux namespace), MLX Metal 0.32.4, using AbstractFramework's mixed Q8/bfloat16 conversion of Wan-AI/Wan2.2-TI2V-5B-Diffusers in the Wan2.2-TI2V-5B family, conversion revision 6875952a110b6bdbcfc00d72b1d89a8e02ab0fc3. Runtime-sensitive attention paths are converted to bfloat16 at load. This is not the publisher's CUDA implementation.
The exact prompt was:
A plain blue ceramic mug remains on a pale wooden table. The camera slowly moves closer. Warm morning light crosses the handle. The mug keeps the same shape and blue color throughout the shot. One continuous shot, no text, no logos, no people.
Download the original MP4. The clip is silent.
| Setting or observation | This attempt |
|---|---|
| Output | 512×320, 49 frames at 24 fps, 2.042 seconds, H.264 |
| Generation | 20 steps, seed 42, guidance 5, flow shift 5, UniPC solver |
| Memory mode | Low RAM, 1 GiB MLX cache limit, streamed VAE decoding |
| Total guarded elapsed | 353.73 seconds, about 5 min 54 s |
| Sampled peak process-tree physical footprint | 14.418 GiB |
| Sampled peak process-tree RSS | 9.648 GiB |
| Additional system swap during this run | 0 GiB observed |
Elapsed time includes process startup, model loading, generation, decoding and saving; it excludes downloads and queue waiting. Loading was not timed separately. This is one cold pipeline observation, with no separate warm-up or repeated timing. Footprint, RSS and Metal allocator memory are different measures; this is not a minimum-RAM claim. The runtime's pre-save generation number is not substituted for the full elapsed time.
We decoded all 49 frames and inspected the complete frame sequence as a contact sheet. The blue mug and warm tabletop are present, but the framing cuts off the rim and part of the mug. Motion is slight and highlights around the handle wobble toward the end. This is a working preview, not an approved product shot. The runtime documentation treats TI2V-5B canvases below 832×480 as functional checks rather than visual-quality settings; this small test should not determine which video model looks best.
Exact prompt, output checksum and inspection notes, runtime metadata and resource measurements preserve this attempt. No HunyuanVideo or H3 generation was run for this guide; no robotics trial was added.
A practical evaluation: the same shot, three attempts
Use a single source image and a short motion brief across candidates. A product shot is useful because identity errors are easy to spot. For example:
A plain ceramic mug remains on a wooden table. The camera slowly moves closer. Morning light crosses the handle. The mug keeps the same shape and color throughout the shot.
Use an image you own and avoid text or logos in the first test. Keep duration and output resolution comparable. Save every attempt rather than selecting one good result from a larger hidden batch.
Review each clip for object identity, motion, camera movement, flicker and the last frame. Then test your actual packaging or product label. Report how many generations produced footage you would keep. If audio is required, score it separately: presence of an audio track does not establish usable speech, timing or sound design.
For timing, include the full generation and file-writing stage. Report model loading separately. For cost, divide the cost of all attempts by the seconds of footage you approved. That measures a useful business result; a nominal price for one generated second does not.
When self-hosting is worth the work
Self-hosting is a reasonable candidate when you need a specific customization, operate a sustained workload, or already maintain compatible GPU infrastructure. Occasional short clips may be easier to produce through a hosted service. Include setup and maintenance time in the comparison.
The MiniMax card describes a context-processing stage, base generation and a higher-resolution regeneration stage. That is a reminder to evaluate the complete intended workflow, not just one downloaded weight file. Pin your model and runtime versions before comparing runs; keep the original source image, prompt and settings with each output.
For a supported creation path, see Teamday Video. This research does not add the listed self-hosted checkpoints to Teamday's model picker. Create a clip in Teamday Video →
For narration, see our Hugging Face TTS samples. For API access instead of self-hosting, compare image and video API providers.
Checkpoint revision and license-metadata snapshots, retrieved 2026-10-31.