Open Video Models 2026: MAGI-2 vs Wan vs LTX-2.5 vs MiniMax H3
We compare MAGI-2, Wan 2.2, LTX-2.5, and MiniMax H3 across Artificial Analysis Elo benchmarks, licensing restrictions, hardware needs, and API costs.
👁 115 reads
In October 2026, four families of open video models stand out: MAGI-2 by Sand.ai, Wan by Alibaba, LTX-2.5 by Lightricks, and MiniMax H3. MiniMax H3 ranks highest on the leaderboard, but its license excludes users in the European Union, the US, and the UK. MAGI-2 offers the cleanest license alongside second-place performance, but it requires 8 GPUs with 80 GB VRAM each. For a standard home workstation with an RTX 4090, Wan 2.2 and LTX-2.5 remain the only practical options. You can find detailed coverage of Sand.ai’s new model in our MAGI-2 Preview guide.
Verdict in Brief (October 11, 2026)
- Best quality with open weights: MiniMax H3 - #2 in image-to-video (1181 Elo). In excluded territories, running weights locally requires separate permission from MiniMax.
- Cleanest license with high quality: MAGI-2 Preview - Apache-2.0, 10th place (1093 Elo), but demands 8 x Hopper 80 GB.
- For a single 24 GB GPU: Wan 2.2 TI2V-5B (Apache-2.0, 720p) or LTX-2.5 in ComfyUI.
- Lowest API cost per minute: MiniMax H3 in text-to-video - $4.80 according to AA. MAGI-2 has no hosted API pricing yet.
Leaderboard: How Open Models Compare
I reviewed both Artificial Analysis leaderboards on October 11, 2026. In the “Weights” column, AA links to a repository whenever a model provides public weights. The latest Wan 3.0 has no such link; Alibaba has only released open weights for the Wan 2.2 series, which is not listed on current AA charts.
| Model | I2V: Rank & Elo | 95% CI | T2V: Rank & Elo | 95% CI |
|---|---|---|---|---|
| MiniMax H3 | 2nd / 1181 | 1173-1189 | 4th / 1137 (768p) | 1128-1146 |
| MAGI-2 Preview | 10th / 1093 | 1086-1100 | 21st / 960 (version 0912, weights unreleased) | 950-970 |
| LTX-2.5 Fast | 23rd / 1038 | 1029-1047 | 23rd / 946 | 936-956 |
| LTX-2.5 Pro | 27th / 1007 | 998-1016 | 24th / 943 | 933-953 |
| Wan 3.0 (proprietary, for comparison) | 7th / 1164 | 1156-1172 | 1st / 1156 | 1147-1165 |
Source: AA-Video-I2V v1.0 and AA-Video-T2V v2.0 (1080p, with audio), retrieved October 11, 2026.
Scale starts at 1,000 points (full bar = 1,200). Key takeaways:
- In image-to-video, MAGI-2 sits distinctly behind H3 (an 88-point gap) and comfortably ahead of LTX-2.5 Fast (by 55 points). The confidence intervals do not overlap, indicating statistically significant differences.
- In text-to-video, MAGI-2 version 0912 (960) and LTX-2.5 Fast (946) have overlapping confidence intervals, making them effectively tied. However, weights for version 0912 are not downloadable.
- Notably, LTX-2.5 Fast outperforms LTX-2.5 Pro across both benchmarks.
Licensing: The Biggest Real-World Hurdle
“Open weights” does not mean unconditional freedom. Here is how the licenses on Hugging Face and GitHub compare as of October 11:
| Model | License | Commercial Use | Regional Restrictions |
|---|---|---|---|
| MAGI-2 Preview | Apache-2.0 | Yes, no revenue caps | None |
| Wan 2.2 (e.g., TI2V-5B) | Apache-2.0 | Yes, no revenue caps | None |
| LTX-2.5 | LTX-2.x Community License (Aug 11, 2026) | Free for entities under $10M annual revenue; paid license required above that threshold | Global; downloading requires account login and terms acceptance |
| MiniMax H3 | MiniMax H3 Community License | Free below $20M annual revenue (written permission above); requires “MiniMax H3” attribution in the product UI | Yes - European Union, United Kingdom, South Korea, and the United States are explicitly excluded |
The H3 license defines “Excluded Territories” as the European Union, the United Kingdom, the Republic of Korea, and the United States. In its FAQ, MiniMax points to ambiguous local video AI regulations, including the EU AI Act, as the rationale. Organizations based in these jurisdictions must apply for an individualized commercial license. However, MiniMax’s hosted API operates globally, meaning you can query H3 via the cloud from anywhere, even if you cannot legally host its weights on your own server. For LTX-2.5, revenue thresholds are calculated across your entire corporate parent group, and distributing fine-tuned checkpoints may require a commercial agreement.
Hardware Requirements
The table below lists only specifications documented in the official README files. Where details are unconfirmed by the developers, it is noted directly.
| Model | Parameters | Hardware According to Developers | ComfyUI Support |
|---|---|---|---|
| MAGI-2 Preview | 114B (MoE, approx. 6B active), ~307 GB weights | 8 x NVIDIA Hopper; at 1080p, pipeline stages do not fit together on a single 80 GB GPU | No |
| Wan 2.2 TI2V-5B | 5B | Min. 24 GB VRAM (e.g., RTX 4090) with offloading; 720p, 24 fps | Yes |
| LTX-2.5 | 22B transformer, 12B Gemma 4 text encoder | Minimum VRAM unspecified; offers int8 and NVFP4 (Blackwell) files alongside FP8 CPU offloading | Yes, official workflow templates |
| MiniMax H3 | 33B transformer (13B AdaLN branches, omitted during inference) | Minimum VRAM unspecified; includes SGLang launch scripts configured for 4 GPUs | Yes, ComfyUI tutorial available |
Sources: model cards for MAGI-2, Wan2.2-TI2V-5B, LTX-2.5, and MiniMax H3 (October 11, 2026). Larger Wan 2.2 variants (such as A14B) were not evaluated for memory consumption.
An RTX 4090 cannot run MAGI-2. Whether H3 or LTX-2.5 can comfortably squeeze into 24 GB without severe bottlenecks remains unstated by their authors, and I have not verified this independently. For a single consumer-grade card, Wan 2.2 TI2V-5B is the only safe bet today. For more guidance on self-hosted setups, check out our guides on Local AI and Wan AI.
API Pricing Comparison
If you lack specialized GPU clusters, open weights are most useful when hosted via third-party APIs. Standard rates per minute from Artificial Analysis tables (default configuration):
| Model | USD per Minute, I2V | USD per Minute, T2V | Notes |
|---|---|---|---|
| MiniMax H3 | $7.80 | $4.80 (768p) | Global MiniMax API; proprietary H3 Max is $4.80 in I2V |
| LTX-2.5 Fast | $7.80 | $7.80 | Lightricks API |
| LTX-2.5 Pro | $10.20 | $10.20 | Lightricks API |
| MAGI-2 Preview | N/A | N/A | AA status: “Coming soon”; unlisted on fal.ai, Replicate, WaveSpeedAI, or OpenRouter |
| Wan 3.0 (proprietary) | $12.00 | $12.00 | Included for comparison |
Full bar represents $12.00. MAGI-2 is excluded from the chart due to lack of public API pricing. The only published hosting estimate suggests $25-$42 per 10-second clip on rented AWS capacity (mushroom.cv blog, August 7, 2026), roughly an order of magnitude more expensive than hosted provider rates. To explore full pricing trends across the industry, read How Much Does a Second of AI Video Cost?.
Feature Comparison
| Feature | MAGI-2 Preview | MiniMax H3 | Wan 2.2 TI2V-5B |
|---|---|---|---|
| Clip Length | Fixed at 10s | 4-15s | 5s per documentation (<9 min runtime on consumer GPUs) |
| Resolution | 1080p (two-stage pipeline) | 768p, 2K via re-generation | 720p, 24 fps |
| Audio | Native generation | Native stereo at 32 kHz | Not specified in model documentation |
| Input Modalities | Text, Text + Image | Text, First/Last Frame, up to 9 images, 3 videos, 3 audio files | Text, Text + Image |
| Prompt Timeline | Yes, structured JSON schema | Descriptive timeline in text prompt | No |
According to its model card, LTX-2.5 supports native multi-shot generation in a single pass, integrated audio generation, and an optional clip-duration predictor.
Which Model Should You Choose?
- If you run a single RTX 4090: Wan 2.2 TI2V-5B (guaranteed to run in 24 GB VRAM) or LTX-2.5 via ComfyUI if you need synchronized audio and multi-shot workflows. MAGI-2 is out of reach.
- If you need commercial deployment with zero licensing friction: MAGI-2 Preview on an 8 x H100/H200 cluster. Its Apache-2.0 license carries no revenue hurdles or regional bans, making it unique among top-10 video models.
- If output fidelity is your priority and weight ownership is secondary: Query MiniMax H3 or the proprietary H3 Max through their cloud API, which delivers higher benchmark scores at lower costs per minute.
- If your enterprise generates over $10M annually: The LTX-2.5 license requires a custom enterprise contract; stick with MAGI-2 or Wan 2.2 to remain on permissive terms.
Frequently Asked Questions
What is the highest-ranked open-weights video model in 2026?
According to Artificial Analysis leaderboards, MiniMax H3 leads the pack, holding #2 overall in image-to-video. MAGI-2 Preview ranks 10th, followed by LTX-2.5 Fast at 23rd.
Can I run MiniMax H3 locally in the US, UK, or EU?
The community license explicitly restricts deployment in “Excluded Territories,” which includes the United States, United Kingdom, European Union, and South Korea. You can access H3 through MiniMax’s global cloud API or contact them for an individualized enterprise exception.
Are model weights available for Wan 3.0?
No open weights have been published for Wan 3.0. Artificial Analysis tracks it as a closed API model, and Alibaba’s official Hugging Face repository (Wan-AI) offers public weights only up to the Wan 2.2 generation.
Which open video model fits inside 24 GB of VRAM?
Wan 2.2 TI2V-5B is explicitly confirmed by its developers to work within 24 GB of VRAM using offloading. MAGI-2 requires an enterprise 8-GPU node and will not boot on an RTX 4090.
Sources and Verification Dates
- Artificial Analysis: Image-to-Video Leaderboard and Text-to-Video Leaderboard (Retrieved Oct 11, 2026)
- MAGI-2 Preview README (Oct 11, 2026)
- MiniMax H3 License and License Q&A (Oct 11, 2026)
- LTX-2.x Community License and LTX-2.5 Model Card (Oct 11, 2026)
- Wan2.2-TI2V-5B Model Card (Oct 11, 2026)
- mushroom.cv: MAGI-2 Hosting Cost Analysis (Aug 7, 2026)
